ESS.XGP[AM,DBL] - www.SailDart.org

perm filename ESS.XGP[AM,DBL] blob sn#415908 filedate 1978-08-07 generic text, type T, neo UTF8
/LMAR=0/XLINE=4/FONT#0=BAXL30/FONT#1=BAXI30/FONT#2=BASB30/FONT#3=BAXS30/FONT#4=GRFX25/FONT#5=FIX20/FONT#6=NGR25/FONT#7=NGR20/FONT#8=BDR66/FONT#9=GACB25/FONT#10=SUP/FONT#11=SUB/FONT#12=METSB
␈↓ α←␈↓␈↓TABLE OF CONTENTS␈↓ 
,vii␈↓


␈↓"β␈↓ α←␈↓α␈↓ ¬]Table of Contents



␈↓ α←␈↓␈↓ βo␈↓αSection ␈↓ 
∞Page␈↓

␈↓ α←␈↓1.␈↓ β∂Introduction
␈↓ α←␈↓␈↓ β∂1-1␈↓ βWContext␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
_  1  
␈↓ α←␈↓␈↓ β∂1-2␈↓ βWTask␈↓ ∧7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
_  2  
␈↓ α←␈↓␈↓ β∂1-3␈↓ βWScope of the problem␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
_  6  
␈↓ α←␈↓␈↓ β∂1-4␈↓ βWMethod␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
_  6  
␈↓ α←␈↓␈↓ β∂1-5␈↓ βWRange of application␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
_  7  
␈↓ α←␈↓␈↓ β∂1-6␈↓ βWA word about natural language␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
_  8  
␈↓ α←␈↓␈↓ β∂1-7␈↓ βWThemes␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
_  8  

␈↓ α←␈↓2.␈↓ β∂Background
␈↓ α←␈↓␈↓ β∂2-1␈↓ βWIntroduction␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  11  
␈↓ α←␈↓␈↓ β∂2-2␈↓ βWComputer consultants␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  11  
␈↓ α←␈↓␈↓ βW2-2-1␈↓ ∧7Mycin␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  12  
␈↓ α←␈↓␈↓ β∂2-3␈↓ βWTeiresias:  System organization␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  15  
␈↓ α←␈↓␈↓ βW2-3-1␈↓ ∧7Performance program:  Knowledge base organization␈↓ 	≠. . . . . . ␈↓ 
	  16  
␈↓ α←␈↓␈↓ βW2-3-2␈↓ ∧7Performance program:  The inference engine␈↓ 	β. . . . . . . . ␈↓ 
	  18  
␈↓ α←␈↓␈↓ βW2-3-3␈↓ ∧7Domain independence and range of application␈↓ 	3. . . . . ␈↓ 
	  20  
␈↓ α←␈↓␈↓ β∂2-4␈↓ βWProduction rules␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  22  
␈↓ α←␈↓␈↓ βW2-4-1␈↓ ∧7Production rules in general␈↓ πC. . . . . . . . . . . . . . . . . . . . ␈↓ 
	  22  
␈↓ α←␈↓␈↓ βW2-4-2␈↓ ∧7Production rules as a knowledge representation␈↓ 	3. . . . . ␈↓ 
	  23  
␈↓ α←␈↓␈↓ βW2-4-3␈↓ ∧7Impact on knowledge organization␈↓ λ#. . . . . . . . . . . . . . ␈↓ 
	  24  
␈↓ α←␈↓␈↓ βW2-4-4␈↓ ∧7Production rules as a high-level language␈↓ 	β. . . . . . . . ␈↓ 
	  25  
␈↓ α←␈↓␈↓ β∂2-5␈↓ βWLevels of knowledge␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  27  

␈↓ α←␈↓3.␈↓ β∂Explanation
␈↓ α←␈↓␈↓ β∂3-1␈↓ βWIntroduction␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  31  
␈↓ α←␈↓␈↓ β∂3-2␈↓ βWBasic assumptions␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  32  
␈↓ α←␈↓␈↓ βW3-2-1␈↓ ∧7Generalities:  Two assumptions␈↓ πs. . . . . . . . . . . . . . . . . ␈↓ 
	  32  
␈↓ α←␈↓␈↓ βW3-2-2␈↓ ∧7Specifics:  How the assumptions were applied␈↓ 	β. . . . . . . . ␈↓ 
	  33  
␈↓ α←␈↓␈↓ β∂3-3␈↓ βWDesign criteria␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  35  
␈↓ α←␈↓␈↓ β∂3-4␈↓ βWBasic ideas␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  36  
␈↓ α←␈↓␈↓ β∂3-5␈↓ βWExplanations for performance validation␈↓ πs. . . . . . . . . . . . . . . . . ␈↓ 
	  37  
␈↓ α←␈↓␈↓ β∂3-6␈↓ βWThe need for an information metric␈↓ πC. . . . . . . . . . . . . . . . . . . . ␈↓ 
	  42  
␈↓ α←␈↓␈↓ β∂3-7␈↓ βWMore sophisticated HOWs␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  43  
␈↓ α←␈↓␈↓ β∂3-8␈↓ βWSpecial purpose responses␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  47  
␈↓ α←␈↓␈↓ β∂3-9␈↓ βWLimitations␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  48  
␈↓ α←␈↓␈↓ βW3-9-1␈↓ ∧7Restriction to a single framework␈↓ πs. . . . . . . . . . . . . . . . . ␈↓ 
	  48  
␈↓ α←␈↓␈↓ βW3-9-2␈↓ ∧7Lack of a general model of explanation␈↓ λS. . . . . . . . . . . ␈↓ 
	  48  
␈↓ α←␈↓␈↓ βW3-9-3␈↓ ∧7Lack of ability to represent control structures␈↓ 	β. . . . . . . . ␈↓ 
	  50  
␈↓ α←␈↓␈↓viii␈↓ λSTABLE OF CONTENTS␈↓

␈↓"β␈↓ α←␈↓␈↓ βW3-9-4␈↓ ∧7Other communication media␈↓ πC. . . . . . . . . . . . . . . . . . . . ␈↓ 
	  52  
␈↓ α←␈↓␈↓ β∂3-10␈↓ βWExplanations for system debugging␈↓ πC. . . . . . . . . . . . . . . . . . . . ␈↓ 
	  52  
␈↓ α←␈↓␈↓ β∂3-11␈↓ βWSUMMARY␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  58  

␈↓ α←␈↓4.␈↓ β∂Knowledge Acquisition:  Overview
␈↓ α←␈↓␈↓ β∂4-1␈↓ βWIntroduction␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  59  
␈↓ α←␈↓␈↓ β∂4-2␈↓ βWPerspective on knowledge acquisition␈↓ πC. . . . . . . . . . . . . . . . . . . . ␈↓ 
	  59  
␈↓ α←␈↓␈↓ β∂4-3␈↓ βWKnowledge acquisition in context␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  60  
␈↓ α←␈↓␈↓ β∂4-4␈↓ βWKnowledge base management␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  61  
␈↓ α←␈↓␈↓ β∂4-5␈↓ βWSystem diagram␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  62  

␈↓ α←␈↓5.␈↓ β∂Knowledge Acquisition I
␈↓ α←␈↓␈↓ β∂5-1␈↓ βWIntroduction␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  65  
␈↓ α←␈↓␈↓ β∂5-2␈↓ βWDebugging example continued␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  66  
␈↓ α←␈↓␈↓ β∂5-3␈↓ βWRule model overview␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  77  
␈↓ α←␈↓␈↓ βW5-3-1␈↓ ∧7Perspective:  Model-based computer vision␈↓ 	β. . . . . . . . ␈↓ 
	  77  
␈↓ α←␈↓␈↓ βW5-3-2␈↓ ∧7Rule models:  Overview␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  79  
␈↓ α←␈↓␈↓ βW5-3-3␈↓ ∧7Rule models:  Example␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  81  
␈↓ α←␈↓␈↓ βW5-3-4␈↓ ∧7Rule models as concept formation␈↓ πs. . . . . . . . . . . . . . . . . ␈↓ 
	  84  
␈↓ α←␈↓␈↓ βW5-3-5␈↓ ∧7Implications␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  85  
␈↓ α←␈↓␈↓ βW5-3-6␈↓ ∧7Character and use of the models␈↓ πs. . . . . . . . . . . . . . . . . ␈↓ 
	  87  
␈↓ α←␈↓␈↓ β∂5-4␈↓ βWHow it all works␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  88  
␈↓ α←␈↓␈↓ βW5-4-1␈↓ ∧7Tracking down the bug␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  88  
␈↓ α←␈↓␈↓ βW5-4-2␈↓ ∧7Deciphering the English text␈↓ πC. . . . . . . . . . . . . . . . . . . . ␈↓ 
	  92  
␈↓ α←␈↓␈↓ βW5-4-3␈↓ ∧7Checking results␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  102  
␈↓ α←␈↓␈↓ βW5-4-4␈↓ ∧7Second guessing␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  105  
␈↓ α←␈↓␈↓ βW5-4-5␈↓ ∧7Final checkout␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  107  
␈↓ α←␈↓␈↓ βW5-4-6␈↓ ∧7Bookkeeping␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  108  
␈↓ α←␈↓␈↓ βW5-4-7␈↓ ∧7Rerunning the consultation␈↓ πC. . . . . . . . . . . . . . . . . . . ␈↓ 	z  109  
␈↓ α←␈↓␈↓ β∂5-5␈↓ βWOther uses for the rule models␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  110  
␈↓ α←␈↓␈↓ βW5-5-1␈↓ ∧7``Knowing what you know'':  Rule models as abstract
␈↓ α←␈↓␈↓ ∧7descriptions of knowledge␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  110  
␈↓ α←␈↓␈↓ βW5-5-2␈↓ ∧7``Knowing what you don't know''␈↓ πs. . . . . . . . . . . . . . . . ␈↓ 	z  110  
␈↓ α←␈↓␈↓ β∂5-6␈↓ βWPersonalized world views␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  112  
␈↓ α←␈↓␈↓ β∂5-7␈↓ βWMore on models, concept formation, and model-based
␈↓ α←␈↓␈↓ βWunderstanding␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  113  
␈↓ α←␈↓␈↓ βW5-7-1␈↓ ∧7Model-based understanding␈↓ πC. . . . . . . . . . . . . . . . . . . ␈↓ 	z  113  
␈↓ α←␈↓␈↓ βW5-7-2␈↓ ∧7Concept formation␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  115  
␈↓ α←␈↓␈↓ βW5-7-3␈↓ ∧7A synthesis␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  116  
␈↓ α←␈↓␈↓ β∂5-8␈↓ βWUnsolved problems and future work␈↓ πC. . . . . . . . . . . . . . . . . . . ␈↓ 	z  118  
␈↓ α←␈↓␈↓ βW5-8-1␈↓ ∧7Minor problems␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  118  
␈↓ α←␈↓␈↓ βW5-8-2␈↓ ∧7Major problems␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  119  
␈↓ α←␈↓␈↓ β∂5-9␈↓ βWSummary␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  123  

␈↓ α←␈↓6.␈↓ β∂Knowledge Acquisition II
␈↓ α←␈↓␈↓ β∂6-1␈↓ βWIntroduction␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  125  
␈↓ α←␈↓␈↓TABLE OF CONTENTS␈↓ 
5ix␈↓

␈↓"β␈↓ α←␈↓␈↓ β∂6-2␈↓ βWKey ideas:   Overview␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  126  
␈↓ α←␈↓␈↓ β∂6-3␈↓ βWThe fundamental problem␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  128  
␈↓ α←␈↓␈↓ β∂6-4␈↓ βWSources of difficulty␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  128  
␈↓ α←␈↓␈↓ β∂6-5␈↓ βWThe solution␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  130  
␈↓ α←␈↓␈↓ β∂6-6␈↓ βWKey ideas:  Comments␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  131  
␈↓ α←␈↓␈↓ βW6-6-1␈↓ ∧7Vocabulary␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  131  
␈↓ α←␈↓␈↓ βW6-6-2␈↓ ∧7Schemata as knowledge representation descriptions␈↓ 	≠. . . . . ␈↓ 	z  131  
␈↓ α←␈↓␈↓ βW6-6-3␈↓ ∧7A ``totally typed'' language␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  133  
␈↓ α←␈↓␈↓ βW6-6-4␈↓ ∧7Knowledge base integrity␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  134  
␈↓ α←␈↓␈↓ βW6-6-5␈↓ ∧7Summary␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  135  
␈↓ α←␈↓␈↓ β∂6-7␈↓ βWAcquiring new values␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  136  
␈↓ α←␈↓␈↓ βW6-7-1␈↓ ∧7Acquisition of a new organism identity␈↓ λS. . . . . . . . . . ␈↓ 	z  136  
␈↓ α←␈↓␈↓ βW6-7-2␈↓ ∧7Acquisition of a new culture site␈↓ πs. . . . . . . . . . . . . . . . ␈↓ 	z  140  
␈↓ α←␈↓␈↓ β∂6-8␈↓ βWKnowledge about representations:  Organization␈↓ λS. . . . . . . . . . ␈↓ 	z  143  
␈↓ α←␈↓␈↓ βW6-8-1␈↓ ∧7The schema hierarchy␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  143  
␈↓ α←␈↓␈↓ βW6-8-2␈↓ ∧7Schema organization␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  145  
␈↓ α←␈↓␈↓ βW6-8-3␈↓ ∧7Slotnames and slotexperts␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  151  
␈↓ α←␈↓␈↓ β∂6-9␈↓ βWKnowledge about representations:  Use␈↓ πs. . . . . . . . . . . . . . . . ␈↓ 	z  154  
␈↓ α←␈↓␈↓ βW6-9-1␈↓ ∧7Schema function:  Acquisition of new instances␈↓ 	3. . . . ␈↓ 	z  154  
␈↓ α←␈↓␈↓ βW6-9-2␈↓ ∧7Where to start in the network␈↓ πC. . . . . . . . . . . . . . . . . . . ␈↓ 	z  160  
␈↓ α←␈↓␈↓ βW6-9-3␈↓ ∧7Schema function:  Access and storage␈↓ λ#. . . . . . . . . . . . . ␈↓ 	z  162  
␈↓ α←␈↓␈↓ β∂6-10␈↓ βWAcquiring a new attribute␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  164  
␈↓ α←␈↓␈↓ βW6-10-1␈↓ ∧7Comments on the trace␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  171  
␈↓ α←␈↓␈↓ β∂6-11␈↓ βWKnowledge about knowledge about representations␈↓ 	β. . . . . . . ␈↓ 	z  173  
␈↓ α←␈↓␈↓ βW6-11-1␈↓ ∧7The SCHEMA-SCHEMA␈↓ πC. . . . . . . . . . . . . . . . . . . ␈↓ 	z  175  
␈↓ α←␈↓␈↓ β∂6-12␈↓ βWBuilding the schema network␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  178  
␈↓ α←␈↓␈↓ βW6-12-1␈↓ ∧7Comments on the trace␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  181  
␈↓ α←␈↓␈↓ β∂6-13␈↓ βWLevels of knowledge␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  183  
␈↓ α←␈↓␈↓ βW6-13-1␈↓ ∧7Level of detail␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  183  
␈↓ α←␈↓␈↓ βW6-13-2␈↓ ∧7Level of generality␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  184  
␈↓ α←␈↓␈↓ βW6-13-3␈↓ ∧7Impact␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  185  
␈↓ α←␈↓␈↓ β∂6-14␈↓ βWLimitations␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  186  
␈↓ α←␈↓␈↓ β∂6-15␈↓ βWFuture work␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  188  
␈↓ α←␈↓␈↓ βW6-15-1␈↓ ∧7Minor extensions␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  188  
␈↓ α←␈↓␈↓ βW6-15-2␈↓ ∧7Major extensions␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  189  
␈↓ α←␈↓␈↓ β∂6-16␈↓ βWSummary␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  189  
␈↓ α←␈↓␈↓ βW6-16-1␈↓ ∧7Review of major concepts␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  189  
␈↓ α←␈↓␈↓ βW6-16-2␈↓ ∧7Current capabilities␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  190  

␈↓ α←␈↓7.␈↓ β∂Strategies
␈↓ α←␈↓␈↓ β∂7-1␈↓ βWIntroduction␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  193  
␈↓ α←␈↓␈↓ β∂7-2␈↓ βWThe main ideas␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  194  
␈↓ α←␈↓␈↓ β∂7-3␈↓ βWWhat is a strategy␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  196  
␈↓ α←␈↓␈↓ βW7-3-1␈↓ ∧7Ill structured problems␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  196  
␈↓ α←␈↓␈↓ βW7-3-2␈↓ ∧7Strategies␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  198  
␈↓ α←␈↓␈↓ βW7-3-3␈↓ ∧7Levels of knowledge␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  201  
␈↓ α←␈↓␈↓x␈↓ λSTABLE OF CONTENTS␈↓

␈↓"β␈↓ α←␈↓␈↓ βW7-3-4␈↓ ∧7Building blocks:  Conceptual primitives, language␈↓ 	≠. . . . . ␈↓ 	z  202  
␈↓ α←␈↓␈↓ β∂7-4␈↓ βWMeta-rules␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  204  
␈↓ α←␈↓␈↓ βW7-4-1␈↓ ∧7Format␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  204  
␈↓ α←␈↓␈↓ βW7-4-2␈↓ ∧7Function␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  206  
␈↓ α←␈↓␈↓ βW7-4-3␈↓ ∧7Implications of meta-rules as a strategy encoding␈↓ 	3. . . . ␈↓ 	z  209  
␈↓ α←␈↓␈↓ βW7-4-4␈↓ ∧7Advanced issues␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  210  
␈↓ α←␈↓␈↓ βW7-4-5␈↓ ∧7Explanation and acquisition␈↓ πC. . . . . . . . . . . . . . . . . . . ␈↓ 	z  214  
␈↓ α←␈↓␈↓ βW7-4-6␈↓ ∧7Limitations of the current implementation, future work␈↓ 	W. . ␈↓ 	z  216  
␈↓ α←␈↓␈↓ β∂7-5␈↓ βWBroader implications␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  218  
␈↓ α←␈↓␈↓ βW7-5-1␈↓ ∧7Reference by name vs. reference by description␈↓ 	3. . . . ␈↓ 	z  218  
␈↓ α←␈↓␈↓ βW7-5-2␈↓ ∧7External descriptors vs. content reference␈↓ λS. . . . . . . . . . ␈↓ 	z  219  
␈↓ α←␈↓␈↓ βW7-5-3␈↓ ∧7Implications of content reference as an invocation
␈↓ α←␈↓␈↓ ∧7mechanism␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  219  
␈↓ α←␈↓␈↓ βW7-5-4␈↓ ∧7Limitations␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  233  
␈↓ α←␈↓␈↓ βW7-5-5␈↓ ∧7Future applications:  Choosing control regimes␈↓ 	3. . . . ␈↓ 	z  237  
␈↓ α←␈↓␈↓ βW7-5-6␈↓ ∧7Review␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  239  
␈↓ α←␈↓␈↓ β∂7-6␈↓ βWA Taxonomy, of sorts␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  241  
␈↓ α←␈↓␈↓ βW7-6-1␈↓ ∧7Generality␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  241  
␈↓ α←␈↓␈↓ βW7-6-2␈↓ ∧7Degree of explicitness␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  242  
␈↓ α←␈↓␈↓ βW7-6-3␈↓ ∧7Knowledge organization␈↓ π∪. . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  243  
␈↓ α←␈↓␈↓ βW7-6-4␈↓ ∧7Character␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  244  
␈↓ α←␈↓␈↓ β∂7-7␈↓ βWLimitations of the general formalism␈↓ πC. . . . . . . . . . . . . . . . . . . ␈↓ 	z  247  
␈↓ α←␈↓␈↓ βW7-7-1␈↓ ∧7Program organization␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  247  
␈↓ α←␈↓␈↓ βW7-7-2␈↓ ∧7Intellectual difficulty␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  247  
␈↓ α←␈↓␈↓ βW7-7-3␈↓ ∧7Dynamic rule creation␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  248  
␈↓ α←␈↓␈↓ βW7-7-4␈↓ ∧7Overhead␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  248  
␈↓ α←␈↓␈↓ β∂7-8␈↓ βWSummary␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  250  

␈↓ α←␈↓8.␈↓ β∂Conclusions
␈↓ α←␈↓␈↓ β∂8-1␈↓ βWIntroduction␈↓ ¬#. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  253  
␈↓ α←␈↓␈↓ β∂8-2␈↓ βWReview of major issues␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  253  
␈↓ α←␈↓␈↓ βW8-2-1␈↓ ∧7Forms of meta-level knowledge␈↓ πs. . . . . . . . . . . . . . . . ␈↓ 	z  253  
␈↓ α←␈↓␈↓ βW8-2-2␈↓ ∧7Explanation␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  254  
␈↓ α←␈↓␈↓ βW8-2-3␈↓ ∧7Knowledge acquisition␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  255  
␈↓ α←␈↓␈↓ βW8-2-4␈↓ ∧7Strategies␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  258  
␈↓ α←␈↓␈↓ β∂8-3␈↓ βWGlobal limitations␈↓ ¬S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  259  
␈↓ α←␈↓␈↓ β∂8-4␈↓ βWThe other themes; some speculations␈↓ πC. . . . . . . . . . . . . . . . . . . ␈↓ 	z  262  
␈↓ α←␈↓␈↓ βW8-4-1␈↓ ∧7Why write a program:  Two views␈↓ λ#. . . . . . . . . . . . . ␈↓ 	z  262  
␈↓ α←␈↓␈↓ β∂8-5␈↓ βWProjections␈↓ ∧s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  275  

␈↓ α←␈↓Author Index␈↓ ∧7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  278  
␈↓ α←␈↓Topic Index␈↓ ∧7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  280  
␈↓ α←␈↓References␈↓ ∧7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  282  
␈↓ α←␈↓␈↓LIST OF PERFORMANCE TRACES␈↓ 
"293␈↓


␈↓"β␈↓ α←␈↓␈↓ β?␈↓ ¬≥␈↓αList of Performance Traces␈↓





␈↓"β␈↓ α←␈↓␈↓ βo␈↓αTrace ␈↓ 
∞Page␈↓

␈↓"β␈↓ α←␈↓␈↓ β∂3-5␈↓ βWExplanations for performance validation␈↓ πs. . . . . . . . . . . . . . . . . ␈↓ 
	  37  
␈↓"β␈↓ α←␈↓␈↓ β∂3-10␈↓ βWExplanations for system debugging␈↓ πC. . . . . . . . . . . . . . . . . . . . ␈↓ 
	  52  
␈↓"β␈↓ α←␈↓␈↓ β∂5-2␈↓ βWDebugging example continued␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 
	  66  
␈↓"β␈↓ α←␈↓␈↓ β∂6-7␈↓ βWAcquiring new values␈↓ εβ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  136  
␈↓"β␈↓ α←␈↓␈↓ β∂6-10␈↓ βWAcquiring a new attribute␈↓ ε3. . . . . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  164  
␈↓"β␈↓ α←␈↓␈↓ β∂6-12␈↓ βWBuilding the schema network␈↓ εc. . . . . . . . . . . . . . . . . . . . . . . . . ␈↓ 	z  178  
␈↓ α←␈↓␈↓␈↓ 
     1␈↓




␈↓"β␈↓ α←␈↓␈↓ ε⊗␈↓αChapter 1



␈↓"β␈↓ α←␈↓α␈↓ ∧R␈↓λINTRODUCTION␈↓α


␈↓"β␈↓ α←␈↓α␈↓ ∧Ithe question;  some goals for the answer









␈↓"β␈↓ α←␈↓␈↓ ¬G␈↓ πQI will tell you the whole truth. 
␈↓"β␈↓ α←␈↓␈↓ 	gline 800

␈↓"β␈↓ α←␈↓␈↓α1-1    CONTEXT␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
creation␈α
and␈α
management␈α
of␈α
large␈α
knowledge␈α
bases␈α
has␈αbecome␈α
a
␈↓ α←␈↓central␈αproblem␈α
of␈αartificial␈α
intelligence␈α(AI)␈α
research.␈α This␈α
is␈αa␈α
result␈αof␈α
two
␈↓ α←␈↓recent␈α⊃trends: ␈α⊂an␈α⊃emphasis␈α⊃on␈α⊂the␈α⊃use␈α⊃of␈α⊂large␈α⊃stores␈α⊃of␈α⊂domain-specific
␈↓ α←␈↓knowledge␈α∃as␈α⊗a␈α∃base␈α⊗for␈α∃high-performance␈α∃programs,␈α⊗and␈α∃a␈α⊗focus␈α∃on
␈↓ α←␈↓problems␈αtaken␈αfrom␈αreal␈αworld␈αsettings.␈α These␈αtrends␈αare␈αmotivated␈αby␈αthe
␈↓ α←␈↓belief␈α∃that␈α∃artificial␈α∃problems␈α∀may,␈α∃in␈α∃the␈α∃long␈α∀run,␈α∃prove␈α∃more␈α∃of␈α∀a
␈↓ α←␈↓diversion␈αthan␈αa␈αbase␈αfor␈αdevelopment␈αand␈αby␈αthe␈αbelief␈αthat␈αthe␈αfield␈αof␈αAI
␈↓ α←␈↓has␈α
progressed␈α∞far␈α
enough␈α∞to␈α
provide␈α
high␈α∞performance␈α
systems␈α∞capable␈α
of
␈↓ α←␈↓solving␈αreal␈αproblems.␈α Both␈αof␈αthese␈αmean␈αan␈αemphasis␈αon␈αthe␈α
accumulation
␈↓ α←␈↓and␈α⊃management␈α∩of␈α⊃large␈α∩collections␈α⊃of␈α⊃knowledge,␈α∩and␈α⊃in␈α∩many␈α⊃systems
␈↓ α←␈↓embodying␈α∀these␈α∀trends␈α∀(e.g.,␈α∀[MACSYMA74],␈α∀[Buchanan71],␈α∀[Finkel74],
␈↓ α←␈↓[Hart75]),␈α_much␈α_time␈α↔has␈α_been␈α_spent␈α↔building␈α_and␈α_maintaining␈α↔such
␈↓ α←␈↓knowledge␈α∩bases.␈α∩ Yet␈α∩there␈α∩has␈α∩been␈α∩little␈α∩discussion␈α∩or␈α∩analysis␈α∪of␈α∩the
␈↓ α←␈↓concomitant␈α⊗problems.␈α⊗ We␈α⊗attempt␈α⊗to␈α⊗define␈α⊗here␈α⊗some␈α⊗of␈α⊗the␈α⊗issues
␈↓ α←␈↓involved␈α∩and␈α⊃explore␈α∩the␈α∩steps␈α⊃taken␈α∩toward␈α∩solving␈α⊃a␈α∩number␈α∩of␈α⊃these
␈↓ α←␈↓problems.␈α
 We␈α∞describe␈α
a␈α∞computer␈α
program␈α
called␈α∞␈↓¬TEIRESIAS␈↓␈↓
1␈↓␈α
that␈α∞has␈α
been

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[1]␈α↔The␈α↔program␈α↔is␈α↔named␈α↔for␈α⊗the␈α↔blind␈α↔seer␈α↔in␈α↔␈↓↓Oedipus␈α↔the␈α⊗King␈↓
␈↓ α←␈↓[Sophocles27],␈α∂and␈α∂selected␈α∂quotes␈α∂from␈α∞the␈α∂play␈α∂are␈α∂scattered␈α∂through␈α∞the
␈↓ α←␈↓text.␈α⊃ As␈α⊃will␈α⊂become␈α⊃clear,␈α⊃the␈α⊂program,␈α⊃like␈α⊃the␈α⊂prophet,␈α⊃has␈α⊃a␈α⊂``higher
␈↓ α←␈↓order''␈α∪of␈α∪knowledge.␈α∀ However,␈α∪this␈α∪rather␈α∀thin␈α∪analogy␈α∪should␈α∀not␈α∪be
␈↓ α←␈↓␈↓2    INTRODUCTION␈↓ 
#1-1␈↓

␈↓"β␈↓ α←␈↓designed and implemented to deal with some of the important issues.

␈↓"β␈↓ α←␈↓␈↓α1-2    TASK␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
fundamental␈α
problem␈α
discussed␈α∞in␈α
the␈α
following␈α
chapters␈α∞is␈α
the
␈↓ α←␈↓creation␈αof␈αa␈αset␈αof␈αtools␈αfor␈αthe␈αconstruction,␈αmaintenance,␈αand␈αuse␈α
of␈αlarge,
␈↓ α←␈↓domain-specific knowledge bases.
␈↓"β␈↓ α←␈↓␈↓ β?Two␈α∩major␈α∩goals␈α∩were␈α∪used␈α∩as␈α∩guidelines␈α∩in␈α∩creating␈α∪those␈α∩tools.
␈↓ α←␈↓First,␈α∂it␈α⊂should␈α∂be␈α⊂possible␈α∂for␈α⊂an␈α∂expert␈α⊂in␈α∂the␈α⊂domain␈α∂of␈α⊂application␈α∂to
␈↓ α←␈↓``educate''␈α~the␈α~performance␈α→program␈α~interactively,␈α~commenting␈α~on␈α→and
␈↓ α←␈↓correcting␈αits␈αbehavior.␈↓
2␈↓␈α ␈αConsider␈αthe␈αtwo␈αalternative␈αapproaches␈αshown␈αin
␈↓ α←␈↓Fig.␈α∪1-1.␈α∀ In␈α∪the␈α∀traditional␈α∪approach,␈α∀the␈α∪behavior␈α∀of␈α∪the␈α∀program␈α∪is
␈↓ α←␈↓interpreted␈α
by␈α
an␈α
assistant␈α
for␈α
the␈αbenefit␈α
of␈α
an␈α
expert␈α
who␈α
knows␈α
little␈αor␈α
no
␈↓ α←␈↓programming.␈α_ The␈α↔expert's␈α_comments␈α↔on␈α_and␈α↔corrections␈α_to␈α↔program
␈↓ α←␈↓behavior␈α∀are␈α∀in␈α∀turn␈α∀interpreted␈α∃by␈α∀the␈α∀assistant␈α∀who␈α∀then␈α∃makes␈α∀the
␈↓ α←␈↓appropriate changes to the program.



␈↓"␈↓ α←␈↓∧␈↓ αg⊂αααααααααα⊃         ⊂αααααααααααα⊃          ⊂ααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧␈↓ αg~  expert  ~  ←ααα   ~ assistant  ~   ←ααα   ~  performance  ~
␈↓"␈↓ α←␈↓∧␈↓ αg~          ~  ααα→   ~            ~   ααα→   ~    program    ~
␈↓"␈↓ α←␈↓∧␈↓ αg%αααααααααα$         %αααααααααααα$          %ααααααααααααααα$

␈↓"␈↓ α←␈↓␈↓ ¬2The traditional approach


␈↓"␈↓ α←␈↓∧␈↓ ∧π⊂αααααααααα⊃               ⊂ααααααααααααα⊃
␈↓"␈↓ α←␈↓∧␈↓ ∧π~          ~  explanation  ~             ~
␈↓"␈↓ α←␈↓∧␈↓ ∧π~  expert  ~  ←αααααααααα  ~ performance ~
␈↓"␈↓ α←␈↓∧␈↓ ∧π~          ~  αααααααααα→  ~   program   ~
␈↓"␈↓ α←␈↓∧␈↓ ∧π~          ~   knowledge   ~             ~
␈↓"␈↓ α←␈↓∧␈↓ ∧π%αααααααααα$   transfer    %ααααααααααααα$

␈↓"␈↓ α←␈↓␈↓ ¬7An alternative approach


␈↓"␈↓ α←␈↓α␈↓ β|Fig. 1-1.    Building high-performance programs.    

␈↓"β␈↓ α←␈↓␈↓ β?The␈α!alternative␈α!approach␈α!puts␈α!the␈α!expert␈α!in␈α!more␈α direct
␈↓ α←␈↓communication␈αwith␈αthe␈αprogram,␈αso␈αthat␈αhe␈αcan␈αdiscover␈αwhat␈αthe␈αprogram
␈↓ α←␈↓is␈α
doing␈α
and␈α
why,␈α
and␈α
can␈α
modify␈α
it␈α
himself␈α
to␈α
produce␈α
the␈αdesired␈α
behavior.
␈↓ α←␈↓It is this alternative toward which we have been working.
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓pursued␈α∪too␈α∪vigorously.␈α∪ Most␈α∪of␈α∀the␈α∪virtue␈α∪in␈α∪having␈α∪a␈α∪name␈α∀for␈α∪the
␈↓ α←␈↓program␈α∂lies␈α⊂in␈α∂the␈α∂convenience␈α⊂it␈α∂offers␈α⊂for␈α∂reference␈α∂to␈α⊂a␈α∂large␈α⊂body␈α∂of
␈↓ α←␈↓code.

␈↓"β␈↓ α←␈↓[2]␈α
``Expert''␈αis␈α
used␈αto␈α
mean␈α
someone␈αexpert␈α
in␈αan␈α
applications␈α
domain␈αbut
␈↓ α←␈↓assumed to be inexperienced in programming.
␈↓ α←␈↓␈↓1-2␈↓ 	bTASK    3␈↓

␈↓"β␈↓ α←␈↓␈↓ β?In␈αthis␈αsituation,␈αthe␈αinteraction␈αbetween␈αthe␈αexpert␈αand␈αthe␈αprogram
␈↓ α←␈↓resembles␈α∂that␈α∞of␈α∂a␈α∂teacher␈α∞who␈α∂continually␈α∂challenges␈α∞a␈α∂student␈α∂with␈α∞new
␈↓ α←␈↓problems␈α⊃to␈α⊃solve␈α⊃and␈α⊃carefully␈α⊃observes␈α⊃the␈α⊃student's␈α∩performance.␈α⊃ The
␈↓ α←␈↓teacher␈α
may␈αinterrupt␈α
the␈αstudent␈α
to␈αrequest␈α
a␈αjustification␈α
of␈αsome␈α
particular
␈↓ α←␈↓step␈α
taken␈α
in␈α
attacking␈α
the␈α
problem␈α
or␈α
he␈α
may␈α
challenge␈α
the␈α
final␈αresult␈α
(both
␈↓ α←␈↓of␈α↔these␈α↔are␈α_information␈α↔transfer␈α↔from␈α↔right␈α_to␈α↔left,␈α↔which␈α_we␈α↔label
␈↓ α←␈↓``explanation'').␈α This␈αprocess␈αmay␈αuncover␈αa␈αfault␈αin␈αthe␈αstudent's␈αknowledge
␈↓ α←␈↓of␈α⊗the␈α⊗subject␈α⊗and␈α∃result␈α⊗in␈α⊗the␈α⊗transfer␈α∃of␈α⊗information␈α⊗to␈α⊗correct␈α∃it
␈↓ α←␈↓(information flow from left to right, ``knowledge transfer'').
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂second␈α∂goal␈α∞was␈α∂to␈α∂make␈α∂it␈α∞possible␈α∂to␈α∂assemble␈α∂and␈α∞maintain
␈↓ α←␈↓large␈α⊂amounts␈α⊃of␈α⊂knowledge.␈α⊂ It␈α⊃is␈α⊂rarely␈α⊂possible␈α⊃to␈α⊂put␈α⊂together␈α⊃a␈α⊂large
␈↓ α←␈↓knowledge␈αbase␈αin␈αone␈αor␈αeven␈αa␈αfew␈αpasses.␈α The␈αprocess␈αis,␈αinstead,␈αone␈αof
␈↓ α←␈↓constant␈α∀trial␈α∀and␈α∀reevaluation,␈α∪an␈α∀incremental␈α∀approach␈α∀to␈α∪competence.
␈↓ α←␈↓When␈α
knowledge␈α
is␈α
accumulated␈α
over␈α
a␈α
long␈α
period␈α
of␈α
time,␈α
the␈α
knowledge
␈↓ α←␈↓base␈αundergoes␈αnumerous␈αchanges.␈α If␈αit␈αis␈αto␈αgrow␈αvery␈αlarge,␈α
making␈αthose
␈↓ α←␈↓changes␈αmust␈αbe␈αa␈αreasonable␈αtask.␈α From␈αthis␈αsimple␈αobservation␈αcomes␈αone
␈↓ α←␈↓major␈α⊃theme␈α⊂of␈α⊃this␈α⊂work: ␈α⊃the␈α⊂search␈α⊃for␈α⊂knowledge␈α⊃representations␈α⊂and
␈↓ α←␈↓system␈αdesigns␈α
that␈αoffer␈α
a␈αhigh␈α
degree␈αof␈α
flexibility␈αin␈α
the␈αface␈α
of␈αchanges.
␈↓ α←␈↓One␈α⊂way␈α⊂to␈α⊂achieve␈α⊂flexibility␈α⊂is␈α⊂to␈α⊂build␈α⊂a␈α⊂separate␈α⊂acquisition␈α∂program
␈↓ α←␈↓specifically␈α↔tailored␈α_to␈α↔the␈α_structure␈α↔of␈α↔the␈α_knowledge␈α↔base.␈α_ A␈α↔more
␈↓ α←␈↓fundamental␈α∪solution␈α∪would␈α∪involve␈α∪designing␈α∪a␈α∪knowledge␈α∪base␈α∪that␈α∩is
␈↓ α←␈↓inherently␈αflexible,␈αone␈αthat␈αwould␈αeasily␈αaccommodate␈αchanges.␈α Elements␈αof
␈↓ α←␈↓both these approaches will be found in the chapters that follow.
␈↓"β␈↓ α←␈↓␈↓ β?Not␈α
surprisingly,␈α
both␈α
goals--forging␈α
a␈α
direct␈α
link␈α
between␈αexpert␈α
and
␈↓ α←␈↓program,␈α∃and␈α∃assembling␈α∃large␈α∃amounts␈α∃of␈α∃knowledge--have␈α∀significant
␈↓ α←␈↓impacts␈αon␈αthe␈α
design␈αof␈αthe␈α
performance␈αprogram.␈α In␈αparticular,␈α
attempting
␈↓ α←␈↓to␈α∃achieve␈α∀both␈α∃of␈α∀them␈α∃simultaneously␈α∀is␈α∃predicated␈α∀on␈α∃an␈α∀important
␈↓ α←␈↓assumption: ␈α∂that␈α∂it␈α∂is␈α∂possible␈α∂to␈α∂distinguish␈α∂between␈α∂basic␈α∂␈↓↓formalism␈↓␈α∂and
␈↓ α←␈↓␈↓↓degree␈α≤of␈α≥expertise␈↓␈α≤or,␈α≥equivalently,␈α≤that␈α≥the␈α≤control␈α≥structure␈α≤and
␈↓ α←␈↓representation␈α∞can␈α∞be␈α
considered␈α∞separately␈α∞from␈α
the␈α∞knowledge␈α∞base.␈↓
3␈↓␈α
The
␈↓ α←␈↓basic␈α⊂control␈α∂structure(s)␈α⊂and␈α⊂representations␈α∂employed␈α⊂in␈α⊂the␈α∂performance
␈↓ α←␈↓program␈αare␈α
assumed␈αto␈α
be␈αestablished␈αand␈α
debugged,␈αand␈α
the␈αfundamental
␈↓ α←␈↓approach␈α
to␈α
the␈α
problem␈α
is␈α
assumed␈α
acceptable.␈α
 The␈α
task␈α
of␈α
the␈αexpert,␈α
then,
␈↓ α←␈↓is␈αto␈αenlarge␈α
the␈αknowledge␈αbase␈αby␈α
adding␈αnew␈αknowledge␈α
to␈αbe␈αused␈αin␈α
one
␈↓ α←␈↓of␈α∩the␈α∩established␈α∩ways.␈α∩ In␈α∩other␈α∩words,␈α∩we␈α∩are␈α∩assuming␈α∩that␈α∩␈↓↓how␈↓␈α⊃the
␈↓ α←␈↓knowledge␈αis␈αto␈αbe␈αused␈αcan␈αbe␈αsettled␈αby␈αthe␈αselection␈αof␈αone␈αor␈αmore␈αof␈αthe
␈↓ α←␈↓available␈αrepresentations␈αand␈αcontrol␈αstructures.␈α The␈αexpert's␈αtask␈αis␈αthus␈αto
␈↓ α←␈↓enlarge ␈↓↓what␈↓ it is the program knows.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α↔is␈α⊗a␈α↔corollary␈α↔assumption␈α⊗in␈α↔the␈α⊗belief␈α↔that␈α↔the␈α⊗control
␈↓ α←␈↓structures␈α∃and␈α∃representations␈α∀are␈α∃comprehensible␈α∃to␈α∀the␈α∃expert␈α∃(at␈α∀the
␈↓ α←␈↓conceptual␈α∞level),␈α
so␈α∞that␈α
he␈α∞can␈α
express␈α∞his␈α
knowledge␈α∞with␈α
them.␈α∞ This␈α
is

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[3]␈α∩Note␈α∪that␈α∩this␈α∩distinction␈α∪is␈α∩not␈α∩specific␈α∪to␈α∩any␈α∩particular␈α∪system␈α∩or
␈↓ α←␈↓knowledge␈α
representation.␈α
 As␈α
long␈α
as␈αit␈α
is␈α
possible␈α
to␈α
make␈α
this␈αdistinction,
␈↓ α←␈↓the general approach used here will remain valid.
␈↓ α←␈↓␈↓4    INTRODUCTION␈↓ 
#1-2␈↓

␈↓"β␈↓ α←␈↓required␈α∪to␈α∀insure␈α∪that␈α∪the␈α∀expert␈α∪understands␈α∪system␈α∀performance␈α∪well
␈↓ α←␈↓enough␈αto␈αknow␈αwhat␈αto␈αcorrect␈αand␈αto␈αassure␈αthat␈αhe␈αknows␈αhow␈αto␈αexpress
␈↓ α←␈↓the␈α∞required␈α∞knowledge.␈α∞ What␈α∞the␈α∞expert␈α∞sees␈α∞and␈α∞wants␈α∞to␈α∞change␈α∂is␈α∞the
␈↓ α←␈↓external␈α∃behavior␈α∃of␈α∃the␈α∃system.␈α∃ Mapping␈α∃from␈α∃the␈α∃desired␈α∃(external)
␈↓ α←␈↓behavior␈α⊃to␈α∩the␈α⊃necessary␈α∩internal␈α⊃modification␈α⊃is␈α∩often␈α⊃quite␈α∩subtle␈α⊃and
␈↓ α←␈↓requires␈αan␈α
intimate␈αunderstanding␈α
of␈αthe␈α
system␈αstructure.␈α
 Part␈αof␈α
the␈α``art
␈↓ α←␈↓of␈αdebugging''␈α
is␈αan␈αunderstanding␈α
of␈αthis␈αmapping.␈α
 We␈αare␈α
thus␈αassuming
␈↓ α←␈↓that␈αthe␈αrepresentation␈αof␈αknowledge␈αand␈αthe␈αmanner␈αin␈αwhich␈αknowledge␈αis
␈↓ α←␈↓used␈αwill␈αbe␈α
sufficiently␈αcomprehensible␈αto␈αthe␈α
expert␈αthat␈αhe␈αcan␈α
understand
␈↓ α←␈↓program behavior in these terms.␈↓
4␈↓
␈↓"β␈↓ α←␈↓␈↓ β?All␈αof␈αthis␈αmeans␈αthat␈αwe␈αwill␈αbe␈αdealing␈αwith␈αperformance␈αprograms
␈↓ α←␈↓having␈α⊂the␈α∂architecture␈α⊂suggested␈α⊂in␈α∂Fig.␈α⊂1-2.␈α⊂ The␈α∂␈↓↓knowledge␈α⊂base␈↓␈α⊂is␈α∂the
␈↓ α←␈↓program's␈α~store␈α~of␈α≠task-specific␈α~knowledge␈α~that␈α~makes␈α≠possible␈α~high
␈↓ α←␈↓performance.␈α The␈α␈↓↓inference␈αengine␈↓␈αis␈αan␈αinterpreter␈αthat␈αuses␈αthe␈αknowledge
␈↓ α←␈↓base to solve the problem at hand.

␈↓"␈↓ α←␈↓∧␈↓ ¬/⊂ααααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧␈↓ ¬/~   ⊂ααααααααααα⊃   ~
␈↓"␈↓ α←␈↓∧␈↓ ¬/~   ~ INFERENCE ~   ~
␈↓"␈↓ α←␈↓∧␈↓ ¬/~   ~  ENGINE   ~   ~
␈↓"␈↓ α←␈↓∧␈↓ ¬/~   %ααααααααααα$   ~
␈↓"␈↓ α←␈↓∧␈↓ ¬/~   ⊂ααααααααααα⊃   ~
␈↓"␈↓ α←␈↓∧␈↓ ¬/~   ~ KNOWLEDGE ~   ~
␈↓"␈↓ α←␈↓∧␈↓ ¬/~   ~   BASE    ~   ~
␈↓"␈↓ α←␈↓∧␈↓ ¬/~   %ααααααααααα$   ~
␈↓"␈↓ α←␈↓∧␈↓ ¬/%ααααααααααααααααααα$


␈↓"␈↓ α←␈↓α␈↓ βBFig. 1-2.    The architecture of the performance program.    

␈↓ α←␈↓The␈α
main␈α
point␈α
here␈αis␈α
the␈α
explicit␈α
division␈αbetween␈α
these␈α
two␈α
parts␈α
of␈αthe
␈↓ α←␈↓program.␈α∞ This␈α∞design␈α∞is␈α∞in␈α
keeping␈α∞with␈α∞the␈α∞assumption␈α∞noted␈α∞above␈α
that
␈↓ α←␈↓the␈α⊂expert's␈α⊂task␈α⊃would␈α⊂be␈α⊂to␈α⊂augment␈α⊃the␈α⊂knowledge␈α⊂base␈α⊂of␈α⊃a␈α⊂program
␈↓ α←␈↓whose␈α∞control␈α∞structure␈α∞(inference␈α
engine)␈α∞is␈α∞assumed␈α∞both␈α∞appropriate␈α
and
␈↓ α←␈↓debugged.␈α If␈αall␈αof␈αthe␈αcontrol␈αstructure␈αinformation␈αis␈αkept␈αin␈αthe␈αinference
␈↓ α←␈↓engine,␈α∀then␈α∀we␈α∀can␈α∀engage␈α∀the␈α∀domain␈α∀expert␈α∀in␈α∀a␈α∀discussion␈α∃of␈α∀the
␈↓ α←␈↓knowledge␈αbase␈αand␈αbe␈αassured␈αthat␈αthe␈αdiscussion␈αwill␈αdeal␈αonly␈αwith␈αissues
␈↓ α←␈↓of␈α
domain-specific␈α
expertise␈α
(rather␈α
than␈α
with␈α
questions␈α
of␈α
programming␈α
and
␈↓ α←␈↓control␈α⊂structures).␈α∂ If␈α⊂all␈α⊂of␈α∂the␈α⊂domain-specific␈α⊂knowledge␈α∂is␈α⊂kept␈α⊂in␈α∂the

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[4]␈αIn␈αterms␈αof␈αa␈αperformance␈αprogram␈αthat␈αuses␈αa␈αrule-based␈αrepresentation
␈↓ α←␈↓of␈αknowledge,␈α
for␈αinstance,␈αthis␈α
means,␈αmore␈αspecifically,␈α
that␈αwe␈α
assume␈αthe
␈↓ α←␈↓expert␈α∞is␈α∞familiar␈α∂with␈α∞the␈α∞fundamental␈α∂structure,␈α∞organization,␈α∞and␈α∂use␈α∞of
␈↓ α←␈↓production␈α⊃rules.␈α⊂ He␈α⊃need␈α⊃only␈α⊂understand␈α⊃them␈α⊂at␈α⊃the␈α⊃conceptual␈α⊂level,
␈↓ α←␈↓since␈α∩part␈α⊃of␈α∩establishing␈α⊃the␈α∩link␈α⊃between␈α∩expert␈α⊃and␈α∩program␈α⊃involves
␈↓ α←␈↓insulating␈αhim␈αfrom␈αdetails␈αof␈α
implementation.␈α The␈αcentral␈αissue␈αis␈αthat␈α
they
␈↓ α←␈↓share a common language of knowledge expression and use.
␈↓ α←␈↓␈↓1-2␈↓ 	bTASK    5␈↓

␈↓"β␈↓ α←␈↓knowledge␈α_base␈α_then␈α_the␈α_program␈α_should␈α_have␈α_a␈α_degree␈α_of␈α_domain
␈↓ α←␈↓independence,␈α
that␈α
is,␈α
it␈α
should␈α
be␈α
possible␈α
to␈α
``unplug''␈α
one␈α
knowledge␈αbase
␈↓ α←␈↓and ``plug in'' another.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α
division␈α∞is␈α
also␈α∞in␈α
keeping␈α∞with␈α
the␈α∞aim␈α
of␈α∞accumulating␈α
large
␈↓ α←␈↓amounts␈α↔of␈α↔knowledge.␈α↔ Experience␈α↔in␈α↔constructing␈α↔large,␈α⊗task-oriented
␈↓ α←␈↓systems␈α[Feigenbaum71]␈αsuggests␈αthat␈αthis␈αseparation␈αmakes␈αaugmentation␈αof
␈↓ α←␈↓program␈α∀performance␈α∀a␈α∀far␈α∀easier␈α∃task␈α∀than␈α∀would␈α∀be␈α∀the␈α∀case␈α∃if␈α∀the
␈↓ α←␈↓distinction were not maintained.
␈↓"β␈↓ α←␈↓␈↓ β?Given␈α
this␈αgeneral␈α
architecture,␈α
we␈αcan␈α
picture␈α
the␈αsituation␈α
in␈αFig.␈α
1-
␈↓ α←␈↓1␈αin␈αmore␈αdetail,␈αviewing␈α␈↓¬TEIRESIAS␈↓␈αas␈αa␈αmeans␈αof␈αestablishing␈αa␈αlink␈αbetween
␈↓ α←␈↓the domain expert and the performance program, Fig. 1-3.


␈↓"␈↓ α←␈↓∧       ⊂αααααααααααααααααααααααααααααααααααααααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧       ~           TEIRESIAS                                  ~
␈↓"␈↓ α←␈↓∧       ~                          ⊂ααααααααααααααααααααααα⊃   ~
␈↓"␈↓ α←␈↓∧E      ~                          ~  PERFORMANCE PROGRAM  ~   ~
␈↓"␈↓ α←␈↓∧       ~    ⊂ααααααααααααα⊃       ~    ⊂ααααααααααα⊃      ~   ~
␈↓"␈↓ α←␈↓∧X    <===== ~ EXPLANATION ~ <=====~==  ~ INFERENCE ~      ~   ~
␈↓"␈↓ α←␈↓∧       ~    ~             ~       ~    ~  ENGINE   ~      ~   ~
␈↓"␈↓ α←␈↓∧P      ~    %ααααααααααααα$       ~    %ααααααααααα$      ~   ~
␈↓"␈↓ α←␈↓∧       ~                          ~    ⊂ααααααααααα⊃      ~   ~
␈↓"␈↓ α←␈↓∧E      ~    ⊂ααααααααααααα⊃       ~    ~ KNOWLEDGE ~      ~   ~
␈↓"␈↓ α←␈↓∧     =====@ ~ KNOWLEDGE   ~ ======~=@  ~   BASE    ~      ~   ~
␈↓"␈↓ α←␈↓∧R      ~    ~ ACQUISITION ~       ~    %ααααααααααα$      ~   ~
␈↓"␈↓ α←␈↓∧       ~    %ααααααααααααα$       ~                       ~   ~
␈↓"␈↓ α←␈↓∧T      ~                          %ααααααααααααααααααααααα$   ~
␈↓"␈↓ α←␈↓∧       ~                                                      ~
␈↓"␈↓ α←␈↓∧       %αααααααααααααααααααααααααααααααααααααααααααααααααααααα$


␈↓"␈↓ α←␈↓α␈↓ β≤Fig. 1-3.    The expert, ␈↓¬TEIRESIAS␈↓α, and the performance program.    

␈↓"β␈↓ α←␈↓␈↓ β?Given␈α⊃the␈α∩range␈α⊃of␈α⊃detailed␈α∩tasks␈α⊃associated␈α⊃with␈α∩knowledge␈α⊃base
␈↓ α←␈↓construction,␈αit␈αis␈αnot␈αalways␈αeasy␈αto␈αput␈αthe␈αexpert␈αin␈αdirect␈αcontact␈αwith␈αthe
␈↓ α←␈↓program␈α␈↓↓and␈↓␈αkeep␈αthe␈αdialog␈αcomprehensible.␈α In␈αresponse,␈αwe␈αhave␈αdevoted
␈↓ α←␈↓extensive␈α
efforts␈α
to␈α
making␈α
the␈αinteraction␈α
as␈α
``high␈α
level''␈α
as␈α
possible.␈α To␈α
the
␈↓ α←␈↓extent␈αthat␈αit␈αis␈αfeasible,␈αfor␈αinstance,␈αquestions␈αfrom␈αthe␈αsystem␈α
are␈αphrased
␈↓ α←␈↓in␈α_terms␈α↔of␈α_the␈α↔manipulation␈α_of␈α↔objects␈α_in␈α↔the␈α_domain␈α↔and␈α_not␈α↔as
␈↓ α←␈↓manipulations␈α
of␈α
program␈α
structures.␈α
 This␈α
helps␈α
``insulate''␈α
the␈α∞expert␈α
from
␈↓ α←␈↓implementation-level details (see especially chapters 6 and 7).
␈↓"β␈↓ α←␈↓␈↓ β?Our␈α
success␈αwith␈α
this␈αproblem␈α
has␈αbeen␈α
varied.␈α Viewed␈α
in␈α
terms␈αof
␈↓ α←␈↓who␈α∪might␈α∪be␈α∪able␈α∪to␈α∪use␈α∪␈↓¬TEIRESIAS␈↓,␈α∪there␈α∪are␈α∪operations␈α∪that␈α∀could␈α∪be
␈↓ α←␈↓performed␈αby␈αsomeone␈αwho␈αhad␈αonly␈αthe␈αbriefest␈αintroduction␈αto␈αthe␈αsystem;
␈↓ α←␈↓others␈α≠require␈α≤more␈α≠extensive␈α≠experience;␈α≤still␈α≠others␈α≤presume␈α≠some
␈↓ α←␈↓programming␈αexperience;␈αand␈αsome␈αrequire␈αinteractions␈α
comprehensible␈αonly
␈↓ α←␈↓to␈α↔the␈α_program's␈α↔author.␈α↔ There␈α_are␈α↔far␈α↔more␈α_of␈α↔the␈α↔first␈α_sort␈α↔and
␈↓ α←␈↓reassuringly␈α
few␈αof␈α
the␈αlast.␈α
 Those␈α
that␈αremain␈α
low␈αlevel␈α
are␈αinvariably␈α
tasks
␈↓ α←␈↓that␈α
are␈α
both␈αconceptually␈α
difficult␈α
and␈αunfamiliar␈α
to␈α
nonprogrammers␈α(e.g.,
␈↓ α←␈↓designing a data structure for a new representation).
␈↓ α←␈↓␈↓6    INTRODUCTION␈↓ 
#1-3␈↓

␈↓"β␈↓ α←␈↓␈↓α1-3    SCOPE OF THE PROBLEM␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Chapters␈α⊗3␈α⊗through␈α⊗7␈α⊗deal␈α⊗with␈α⊗four␈α⊗problems␈α⊗encountered␈α⊗in
␈↓ α←␈↓attempting␈α
to␈αreach␈α
the␈α
pair␈αof␈α
goals␈α
outlined␈αabove.␈α
 Each␈α
of␈αthese␈α
problems
␈↓ α←␈↓supplied␈α∞a␈α∞major␈α∞topic␈α∞of␈α∞investigation␈α∞and␈α∞was␈α∞considered␈α∞individually␈α
in
␈↓ α←␈↓broadly␈α
applicable␈α
terms.␈α
 Each␈α
is␈αcharacterized␈α
below␈α
by␈α
a␈α
description␈αof␈α
the
␈↓ α←␈↓problem␈α⊃and␈α∩an␈α⊃indication␈α∩of␈α⊃an␈α⊃acceptable␈α∩solution.␈α⊃ The␈α∩chapters␈α⊃that
␈↓ α←␈↓follow␈α
describe␈αthe␈α
attempts␈α
to␈αsolve␈α
these␈α
problems␈αand␈α
explore␈α
the␈αextent␈α
to
␈↓ α←␈↓which an acceptable level of performance has been achieved.
␈↓"β␈↓ α←␈↓␈↓ β?Chapter␈α⊃3␈α⊃discusses␈α⊃efforts␈α∩to␈α⊃enable␈α⊃the␈α⊃performance␈α∩program␈α⊃to
␈↓ α←␈↓explain␈α
its␈αactions.␈α
 The␈αfundamental␈α
goal␈α
was␈αto␈α
design␈αa␈α
facility␈αthat␈α
would
␈↓ α←␈↓allow␈α∞the␈α∂program␈α∞to␈α∞explain␈α∂itself␈α∞to␈α∞a␈α∂wide-ranging␈α∞audience␈α∂that␈α∞might
␈↓ α←␈↓include: ␈α⊂an␈α⊂expert␈α⊂who␈α∂wanted␈α⊂to␈α⊂debug␈α⊂its␈α∂knowledge␈α⊂base,␈α⊂a␈α⊂user␈α∂who
␈↓ α←␈↓requested␈α⊂its␈α⊂services,␈α∂and␈α⊂a␈α⊂student␈α⊂with␈α∂minimal␈α⊂experience␈α⊂in␈α⊂the␈α∂field
␈↓ α←␈↓who wanted to learn from it.
␈↓"β␈↓ α←␈↓␈↓ β?Chapter␈α4␈α
provides␈αa␈αbrief␈α
overview␈αof␈α
our␈αperspective␈αon␈α
knowledge
␈↓ α←␈↓acquisition,␈α
while␈αchapter␈α
5␈αdescribes␈α
techniques␈αthat␈α
make␈αit␈α
possible␈αfor␈α
the
␈↓ α←␈↓expert␈α
to␈αsupply␈α
new␈α
inference␈αrules,␈α
using␈α
a␈αrestricted␈α
subset␈α
of␈αthe␈α
language
␈↓ α←␈↓and␈α⊂vocabulary␈α⊃natural␈α⊂to␈α⊃the␈α⊂domain.␈α⊃ It␈α⊂illustrates␈α⊃the␈α⊂utility␈α⊃of␈α⊂setting
␈↓ α←␈↓knowledge␈αacquisition␈α
in␈αthe␈α
context␈αof␈αshortcomings␈α
in␈αthe␈α
knowledge␈αbase
␈↓ α←␈↓and␈α
describes␈α
how␈α
the␈α
program␈α
forms␈α
expectations␈α
during␈α
its␈α
interaction␈α
with
␈↓ α←␈↓the␈αexpert.␈α The␈αbasic␈αgoal␈αhere␈αwas␈αto␈αmake␈αthe␈αknowledge␈αtransfer␈αprocess
␈↓ α←␈↓both␈α⊂easy␈α⊂enough␈α∂and␈α⊂``intelligent''␈α⊂enough␈α⊂so␈α∂that␈α⊂the␈α⊂expert␈α⊂alone␈α∂could
␈↓ α←␈↓make significant additions to the knowledge base.
␈↓"β␈↓ α←␈↓␈↓ β?Chapter␈α
6␈α
discusses␈α
the␈α
acquisition␈α
of␈α
new␈α
conceptual␈α
primitives.␈α
This
␈↓ α←␈↓chapter␈α∞views␈α∞the␈α∞process␈α∞in␈α∞terms␈α
of␈α∞a␈α∞knowledge␈α∞base␈α∞and␈α∞data␈α
structure
␈↓ α←␈↓management␈α⊗task␈α↔and␈α⊗describes␈α⊗the␈α↔techniques␈α⊗developed␈α↔for␈α⊗effective
␈↓ α←␈↓performance␈α∩when␈α∩data␈α∩structures␈α∩and␈α∩representations␈α∩are␈α∩uncomplicated.
␈↓ α←␈↓More␈α⊂generally,␈α⊂it␈α⊂explores␈α⊂the␈α⊂use␈α⊂of␈α⊂meta-level␈α⊂knowledge␈α⊂as␈α⊂a␈α⊂tool␈α⊂for
␈↓ α←␈↓knowledge␈αbase␈αmanagement.␈α The␈αgoal␈αhere␈αwas␈αto␈αmake␈αit␈αpossible␈αfor␈αthe
␈↓ α←␈↓expert to build an entire knowledge base from scratch.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈αchapter␈α7␈αexamines␈αthe␈αproblem␈αof␈αrepresentation␈αand␈αuse␈αof
␈↓ α←␈↓strategies␈αthat␈αenable␈αa␈αprogram␈αto␈αmake␈αmore␈αefficient␈αuse␈αof␈αits␈αknowledge
␈↓ α←␈↓base.␈α∩ It␈α∪also␈α∩explores␈α∩the␈α∪larger␈α∩question␈α∩of␈α∪meta-level␈α∩knowledge␈α∪as␈α∩a
␈↓ α←␈↓framework␈α∂for␈α∂the␈α∂organization␈α∂and␈α∂expression␈α∂of␈α∂strategies␈α∂and␈α∂examines
␈↓ α←␈↓issues of implementation and efficiency.

␈↓"β␈↓ α←␈↓␈↓α1-4    METHOD␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αcentral␈αtheme␈αof␈αthis␈αwork␈αis␈αthe␈αexploration␈αand␈αuse␈αof␈αwhat␈α1f
␈↓ α←␈↓have␈α
labeled␈α␈↓↓meta-level␈α
knowledge␈↓.␈α This␈α
takes␈αseveral␈α
different␈αforms␈α
as␈αits
␈↓ α←␈↓use␈α∂is␈α⊂explored,␈α∂but␈α∂can␈α⊂be␈α∂summed␈α⊂up␈α∂as␈α∂``knowing␈α⊂what␈α∂you␈α⊂know.'' ␈α∂It
␈↓ α←␈↓makes␈αpossible␈α
a␈αsystem␈α
with␈αboth␈αthe␈α
capacity␈αto␈α
use␈αits␈α
knowledge␈αdirectly
␈↓ α←␈↓and the ability to examine it, abstract it, and direct its application.
␈↓"β␈↓ α←␈↓␈↓ β?To␈α∩see␈α∩how␈α∩this␈α∩might␈α∩be␈α∩done,␈α∩recall␈α∩that␈α∩one␈α∩of␈α∪the␈α∩principal
␈↓ α←␈↓problems␈αof␈αAI␈αis␈αthe␈αquestion␈α
of␈αrepresentation␈αand␈αuse␈αof␈αknowledge␈α
about
␈↓ α←␈↓the␈α∂world.␈α∂ Much␈α∂progress␈α∂has␈α∂been␈α∂made␈α∂and␈α∂numerous␈α∂techniques␈α∂have
␈↓ α←␈↓␈↓1-4␈↓ 	7METHOD    7␈↓

␈↓"β␈↓ α←␈↓been␈αdeveloped.␈α One␈αway␈αto␈αview␈αwhat␈αwe␈αhave␈αdone␈αis␈αto␈αimagine␈αturning
␈↓ α←␈↓this␈α
idea␈α
of␈α
knowledge␈α
representation␈α
in␈αon␈α
itself,␈α
using␈α
some␈α
of␈α
these␈αsame
␈↓ α←␈↓techniques␈αto␈αdescribe␈αparts␈αof␈αthe␈αprogram␈αitself.␈α Thus␈αwe␈αhave␈αa␈αprogram
␈↓ α←␈↓containing␈α⊃both␈α∩object-level␈α⊃representations␈α∩describing␈α⊃the␈α∩external␈α⊃world
␈↓ α←␈↓and␈α%meta-level␈α%representations␈α&describing␈α%the␈α%internal␈α&world␈α%of
␈↓ α←␈↓representations.
␈↓"β␈↓ α←␈↓␈↓ β?In␈αthe␈αgeneral␈αcontext␈αof␈αbuilding␈αa␈αlarge␈αknowledge␈αbase,␈αmeta-level
␈↓ α←␈↓knowledge␈α∪has␈α∪been␈α∪used␈α∪as␈α∩a␈α∪tool␈α∪for␈α∪the␈α∪management␈α∪of␈α∩object-level
␈↓ α←␈↓knowledge.␈α∂ We␈α∂report␈α∂on␈α∂the␈α∂capabilities␈α∂made␈α∂possible␈α∂by␈α∂this␈α∂approach
␈↓ α←␈↓and␈αdocument␈αcases␈α
where␈αits␈αabsence␈α
has␈αresulted␈αin␈α
significant␈αdifficulties.
␈↓ α←␈↓Chapter␈α∃3␈α∀describes␈α∃the␈α∀explanation␈α∃system,␈α∀which␈α∃involves␈α∃giving␈α∀the
␈↓ α←␈↓performance␈α!program␈α!a␈α!model␈α!of␈α!its␈α!control␈α!structures␈α"and␈α!an
␈↓ α←␈↓``understanding''␈α∃of␈α⊗its␈α∃representations;␈α⊗chapter␈α∃5␈α⊗documents␈α∃the␈α⊗use␈α∃of
␈↓ α←␈↓abstracted␈α_models␈α↔of␈α_knowledge␈α_as␈α↔a␈α_guide␈α↔to␈α_acquisition;␈α_chapter␈α↔6
␈↓ α←␈↓demonstrates␈α∀the␈α∪utility␈α∀of␈α∀describing␈α∪to␈α∀the␈α∀system␈α∪the␈α∀structure␈α∀of␈α∪its
␈↓ α←␈↓knowledge␈α∂base;␈α∂and␈α∂chapter␈α∂7␈α⊂describes␈α∂the␈α∂use␈α∂of␈α∂strategies␈α⊂that␈α∂contain
␈↓ α←␈↓knowledge about the use of knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?Keep␈α∂in␈α∞mind␈α∂that␈α∂meta-level␈α∞knowledge␈α∂does␈α∂not␈α∞refer␈α∂to␈α∂a␈α∞single
␈↓ α←␈↓entity␈α
but␈α
is␈α
used␈α
as␈α
a␈αgeneric␈α
term␈α
for␈α
several␈α
different␈α
kinds␈αof␈α
information.
␈↓ α←␈↓Each of these chapters explores one or more manifestations of it.

␈↓"β␈↓ α←␈↓␈↓α1-5    RANGE OF APPLICATION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?We␈α
noted␈α
above␈α
that␈α
it␈α
is␈α∞the␈α
expert's␈α
task␈α
to␈α
enlarge␈α
what␈α
it␈α∞is␈α
the
␈↓ α←␈↓program␈α∀knows,␈α∃working␈α∀within␈α∀the␈α∃existing␈α∀set␈α∀of␈α∃representations␈α∀and
␈↓ α←␈↓control␈α⊂structures␈α⊂to␈α⊂correct␈α⊂shortcomings␈α⊂he␈α⊂finds␈α⊂in␈α⊂the␈α⊃knowledge␈α⊂base.
␈↓ α←␈↓Within␈α⊂this␈α∂framework␈α⊂we␈α⊂can␈α∂imagine␈α⊂a␈α∂range␈α⊂of␈α⊂different␈α∂shortcomings
␈↓ α←␈↓that he might uncover:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?ignorance--some piece of knowledge is missing,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?stupidity--some piece of knowledge is incorrect,
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?incompetence--the␈α_current␈α↔set␈α_of␈α↔conceptual␈α_primitives␈α↔is
␈↓ α←␈↓␈↓ βoincapable of expressing a needed piece of knowledge, and
␈↓"β␈↓ α←␈↓␈↓ ββ(d)␈↓ β?formalism␈α
bug--the␈α
control␈α
structure␈α
has␈α
a␈α
bug,␈α
or␈α
the␈α
set␈α
of
␈↓ α←␈↓␈↓ βoavailable representations is inadequate.

␈↓ α←␈↓From␈α∃these␈α∃we␈α∃can␈α∃get␈α∃a␈α∃feeling␈α∃for␈α∃the␈α∃range␈α∃of␈α∃application␈α⊗of␈α∃the
␈↓ α←␈↓techniques␈α⊃described␈α∩in␈α⊃subsequent␈α⊃chapters.␈α∩ We␈α⊃will␈α⊃be␈α∩concerned␈α⊃here
␈↓ α←␈↓with␈α∞the␈α∂first␈α∞three␈α∞(although␈α∂there␈α∞is␈α∞an␈α∂important␈α∞class␈α∂of␈α∞incompetence-
␈↓ α←␈↓type␈α∞errors␈α
that␈α∞cannot␈α∞be␈α
handled;␈α∞see␈α
chapter␈α∞6)␈α∞and␈α
make␈α∞no␈α∞attempt␈α
to
␈↓ α←␈↓deal␈α
with␈α
the␈α
last.␈α This␈α
means␈α
that␈α
the␈αtools␈α
provided␈α
are␈α
capable␈αof␈α
making
␈↓ α←␈↓extensive␈αchanges␈αto␈αthe␈αknowledge␈αbase␈αbut␈αnone␈αat␈αall␈αto␈αthe␈α
basic␈αcontrol
␈↓ α←␈↓structures.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈α∩the␈α∩expert␈α⊃is␈α∩constrained␈α∩to␈α∩work␈α⊃with␈α∩the␈α∩available␈α∩set␈α⊃of
␈↓ α←␈↓representations␈αand␈αcontrol␈αstructures,␈αour␈α
approach␈αis␈αalso␈αlimited␈αto␈α
dealing
␈↓ α←␈↓with␈α∀knowledge␈α∀in␈α∀the␈α∀application␈α∀domain␈α∀that␈α∀can␈α∀be␈α∀formalized␈α∪and
␈↓ α←␈↓␈↓8    INTRODUCTION␈↓ 
#1-5␈↓

␈↓"β␈↓ α←␈↓expressed␈α⊂within␈α⊂the␈α⊂range␈α⊂of␈α⊂available␈α⊂techniques.␈α⊂ This␈α⊂is␈α⊂a␈α⊂substantial
␈↓ α←␈↓assumption,␈α
since␈α
knowledge␈α
in␈α
some␈α
domains␈α
is␈α
ill-specified␈α
and␈α
it␈αis␈α
unclear
␈↓ α←␈↓what␈α∩even␈α⊃the␈α∩basic␈α⊃conceptual␈α∩primitives␈α⊃should␈α∩be.␈α⊃ In␈α∩other␈α⊃domains,
␈↓ α←␈↓processes␈α∃may␈α∀be␈α∃so␈α∃well␈α∀understood␈α∃that␈α∀definitive␈α∃algorithms␈α∃can␈α∀be
␈↓ α←␈↓specified,␈α∪eliminating␈α∪the␈α∪need␈α∪to␈α∪accumulate␈α∪large␈α∪amounts␈α∪of␈α∪informal
␈↓ α←␈↓knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?Our␈α∞approach␈α
to␈α∞knowledge␈α
acquisition␈α∞is␈α
geared␈α∞to␈α∞domains␈α
whose
␈↓ α←␈↓level␈α∩of␈α∩formalization␈α∩falls␈α∪somewhere␈α∩between␈α∩these␈α∩two␈α∪extremes.␈α∩ The
␈↓ α←␈↓``vocabulary''␈α∞of␈α∞conceptual␈α∞primitives␈α∞should␈α∞be␈α∞established,␈α∞but␈α∞knowledge
␈↓ α←␈↓should␈αstill␈αbe␈αincomplete␈αenough␈αthat␈αproblem␈αsolving␈αis␈αa␈αheuristic␈αprocess.
␈↓ α←␈↓The␈α∂knowledge␈α∂should␈α∂also␈α∂be␈α∂decomposable␈α∂into␈α∂small,␈α∂modular␈α∞``chunks''
␈↓ α←␈↓that␈αcan␈αbe␈αexpressed␈αwith␈αa␈αsimple␈αsyntax.␈α The␈αlatter␈αimplies␈αthat␈αboth␈αthe
␈↓ α←␈↓number␈α⊃of␈α⊃interacting␈α⊃factors␈α⊃and␈α⊃the␈α⊃complexity␈α⊃of␈α⊃their␈α⊃interaction␈α⊂are
␈↓ α←␈↓limited.␈α For␈αa␈αrange␈αof␈αtasks,␈αknowledge␈αexpressed␈αin␈αproduction␈αrules␈αmeets
␈↓ α←␈↓both constraints.

␈↓"β␈↓ α←␈↓␈↓α1-6    A WORD ABOUT NATURAL LANGUAGE␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Natural␈α∞language␈α∂has␈α∞not␈α∂been␈α∞a␈α∞major␈α∂focus␈α∞of␈α∂this␈α∞work;␈α∂for␈α∞the
␈↓ α←␈↓most␈αpart,␈αwe␈αhave␈αused␈αthe␈αsimplest␈αtechniques␈αthat␈αwould␈αsupport␈αthe␈α
level
␈↓ α←␈↓of␈α⊃performance␈α∩required.␈α⊃ All␈α∩questions␈α⊃and␈α∩responses␈α⊃from␈α∩␈↓¬TEIRESIAS␈↓␈α⊃are
␈↓ α←␈↓either␈α
pre-formed␈α
or␈α
based␈α∞on␈α
a␈α
simple␈α
template␈α
completion␈α∞mechanism␈α
(as
␈↓ α←␈↓evidenced␈α
by␈α
the␈α
appearance␈α
of␈α∞phrases␈α
like␈α
``a␈α
area'').␈α
 Responses␈α∞from␈α
the
␈↓ α←␈↓user␈α∪are␈α∩of␈α∪three␈α∪general␈α∩types: ␈α∪single-token␈α∩answers␈α∪to␈α∪multiple␈α∩choice
␈↓ α←␈↓questions,␈α
strings␈α
belonging␈αto␈α
a␈α
synthetic␈αlanguage␈α
with␈α
a␈α
formal␈αgrammar,
␈↓ α←␈↓and␈αheavily␈αstylized␈αnatural␈αlanguage␈αsentences␈αusing␈αa␈αrestricted␈αvocabulary
␈↓ α←␈↓(examples␈α∩of␈α∩all␈α∩of␈α∩these␈α∩are␈α⊃seen␈α∩in␈α∩subsequent␈α∩chapters).␈α∩ The␈α∩first␈α⊃is
␈↓ α←␈↓handled␈α∪in␈α∪the␈α∪obvious␈α∪way,␈α∪the␈α∪second␈α∪relies␈α∪on␈α∪a␈α∪simple␈α∀parser␈α∪that
␈↓ α←␈↓matches␈α
user␈α∞input␈α
against␈α∞a␈α
BNF␈α∞specification␈α
of␈α∞valid␈α
responses,␈α∞and␈α
the
␈↓ α←␈↓last relies on straightforward keyword analysis.
␈↓"β␈↓ α←␈↓␈↓ β?This␈αapproach␈αhas␈αserved␈αthus␈αfar␈αto␈αkeep␈αthe␈αinteraction␈αacceptably
␈↓ α←␈↓``natural,''␈αwithout␈αunreasonable␈αprocessing␈αoverhead.␈α It␈αappears␈αto␈αbe␈α
viable
␈↓ α←␈↓where␈α⊂unrestricted␈α⊂dialog␈α⊂is␈α⊂not␈α⊂the␈α∂goal␈α⊂and␈α⊂for␈α⊂domains␈α⊂where␈α⊂there␈α∂is
␈↓ α←␈↓available␈α∂a␈α∞semiformal␈α∂technical␈α∂language␈α∞with␈α∂a␈α∞low␈α∂degree␈α∂of␈α∞ambiguity.
␈↓ α←␈↓Since,␈α∩in␈α∩our␈α∩experience,␈α∪technical␈α∩interchange␈α∩in␈α∩such␈α∩domains␈α∪is␈α∩often
␈↓ α←␈↓ungrammatical␈α∩(relying␈α∩instead␈α∩on␈α∩technical␈α∩terms␈α∩to␈α∩convey␈α∩meaning),␈α∩a
␈↓ α←␈↓heavily grammar-based approach might not have fared well in any case.

␈↓"β␈↓ α←␈↓␈↓α1-7    THEMES␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈α
are␈α
at␈α∞least␈α
two␈α
different,␈α
completely␈α∞orthogonal␈α
organizations
␈↓ α←␈↓of␈α
the␈α
ideas␈α
presented␈α
here.␈α
 The␈α
first␈α
is␈α
suggested␈α
by␈α
the␈α
table␈α
of␈αcontents,
␈↓ α←␈↓which indicates chapters dealing with four basic tasks:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?explanation,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?acquisition of new inference rules,
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?acquisition of new conceptual primitives, and
␈↓"β␈↓ α←␈↓␈↓ ββ(d)␈↓ β?encoding, organization, and use of strategies.
␈↓ α←␈↓␈↓1-7␈↓ 	:THEMES    9␈↓

␈↓"β␈↓ α←␈↓All␈α⊃of␈α⊂these␈α⊃employ␈α⊃techniques␈α⊂based␈α⊃on␈α⊂different␈α⊃varieties␈α⊃of␈α⊂meta-level
␈↓ α←␈↓knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αsecond␈αorganization␈αof␈αthe␈αmaterial␈αis␈αsuggested␈αby␈αthe␈αcollection
␈↓ α←␈↓of␈α∞catch␈α∂phrases␈α∞below.␈α∂ These␈α∞represent␈α∞themes␈α∂that␈α∞recur␈α∂throughout␈α∞the
␈↓ α←␈↓remainder␈α∩of␈α∩this␈α∩work.␈α∩ It␈α∩will␈α∩be␈α∩useful␈α∩to␈α∩keep␈α∩these␈α∩in␈α∩mind␈α∩as␈α∩an
␈↓ α←␈↓alternative␈αset␈αof␈αissues␈αaddressed␈αby␈αthe␈αwork.␈α They␈αwill␈αbe␈αrevisited␈αin␈αthe
␈↓ α←␈↓final␈αchapter␈α
to␈αsee␈αhow␈α
close␈αwe␈α
have␈αcome␈αto␈α
some␈αof␈α
the␈αideals␈αthey␈α
imply.
␈↓ α←␈↓They␈α∪are␈α∪purposely␈α∪oversimplified␈α∪here␈α∪for␈α∪the␈α∪sake␈α∪of␈α∪clarity␈α∪and␈α∩are
␈↓ α←␈↓intended␈α
only␈αto␈α
be␈αsuggestive,␈α
conveying␈αby␈α
keywords␈αand␈α
phrases␈α
some␈αof
␈↓ α←␈↓the character of the work that follows.

␈↓"β␈↓ α←␈↓␈↓ ββ(1)␈↓ β?Task-specific high-level languages make code easier to read.

␈↓"β␈↓ α←␈↓␈↓ ββ(2)␈↓ β?Knowledge in programs should be explicit and accessible.

␈↓"β␈↓ α←␈↓␈↓ ββ(3)␈↓ β?Programs can be self-understanding.

␈↓"β␈↓ α←␈↓␈↓ ββ(4)␈↓ β?Programs␈α∂can␈α∂have␈α∂access␈α∞to␈α∂and␈α∂an␈α∂understanding␈α∂of␈α∞their
␈↓ α←␈↓␈↓ β?own representations.

␈↓"β␈↓ α←␈↓␈↓ ββ(5)␈↓ β?Programs can have some grasp on their own complexity.

␈↓"β␈↓ α←␈↓␈↓ ββ(6)␈↓ β?Programs can be self-adjusting.

␈↓"β␈↓ α←␈↓␈↓ ββ(7)␈↓ β?Representations␈α
can␈α
usefully␈α
be␈α
more␈α
than␈α
a␈α∞densely␈α
encoded
␈↓ α←␈↓␈↓ β?string of bits.
␈↓ α←␈↓␈↓␈↓ 
⊃    11␈↓




␈↓"β␈↓ α←␈↓␈↓ ε∃␈↓αChapter 2



␈↓"β␈↓ α←␈↓α␈↓ ∧p␈↓λBACKGROUND␈↓α


␈↓"β␈↓ α←␈↓α␈↓ αdTEIRESIAS, computer consultants, production rules and some vocabulary









␈↓"β␈↓ α←␈↓␈↓ ¬G␈↓ λ
It vexes me what ails him. 
␈↓"β␈↓ α←␈↓␈↓ 	wline 74

␈↓"β␈↓ α←␈↓␈↓α2-1    INTRODUCTION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
first␈α
part␈α
of␈α
this␈α∞chapter␈α
provides␈α
a␈α
brief␈α
overview␈α∞of␈α
␈↓¬TEIRESIAS␈↓
␈↓ α←␈↓and␈α⊃the␈α⊃sort␈α⊃of␈α⊃performance␈α∩programs␈α⊃it␈α⊃is␈α⊃designed␈α⊃to␈α⊃help␈α∩build.␈α⊃ We
␈↓ α←␈↓describe␈α
the␈α
knowledge␈α
representations␈αused,␈α
review␈α
the␈α
control␈αstructure,␈α
and
␈↓ α←␈↓introduce␈α∀several␈α∀concepts␈α∀and␈α∃terms␈α∀that␈α∀will␈α∀be␈α∀useful␈α∃vocabulary␈α∀in
␈↓ α←␈↓subsequent chapters.
␈↓"β␈↓ α←␈↓␈↓ β?Section␈α∃2-4␈α∃explores␈α∃some␈α∃general␈α∃ideas␈α∃about␈α⊗production␈α∃rules,
␈↓ α←␈↓showing␈αhow␈αthey␈αhave␈αbeen␈α
used␈αto␈αdevelop␈αwhat␈αamounts␈αto␈α
a␈αhigh-level
␈↓ α←␈↓language␈αand␈α
indicating␈αhow␈α
this␈αlanguage␈αforms␈α
the␈αbasis␈α
for␈αmany␈α
of␈αthe
␈↓ α←␈↓capabilities discussed later.
␈↓"β␈↓ α←␈↓␈↓ β?Section␈α⊂2-5␈α⊃considers␈α⊂briefly␈α⊃the␈α⊂problem␈α⊃of␈α⊂high␈α⊃performance␈α⊂vs.
␈↓ α←␈↓generality␈αand␈αexamines␈αthe␈αwork␈αwe␈αhave␈αdone␈αin␈αthat␈αlight,␈αconsidering␈αin
␈↓ α←␈↓particular the benefits of a hierarchical layering of types of knowledge.

␈↓"β␈↓ α←␈↓␈↓α2-2    COMPUTER CONSULTANTS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃recent␈α∩growth␈α⊃of␈α⊃interest␈α∩in␈α⊃the␈α⊃class␈α∩of␈α⊃programs␈α∩known␈α⊃as
␈↓ α←␈↓computer␈α
consultants␈αcan␈α
be␈αseen␈α
as␈αa␈α
logical␈αconsequence␈α
of␈αthe␈α
two␈αtrends
␈↓ α←␈↓noted␈α→in␈α→chapter␈α→1--an␈α→emphasis␈α→on␈α→large␈α→stores␈α→of␈α→domain-specific
␈↓ α←␈↓knowledge␈αand␈αthe␈αconcentration␈αon␈αproblems␈αtaken␈αfrom␈αreal␈αworld␈αsettings.
␈↓ α←␈↓These␈α∂programs␈α∞are␈α∂intended␈α∂to␈α∞provide␈α∂expert-level␈α∞advice␈α∂on␈α∂a␈α∞difficult
␈↓ α←␈↓cognitive␈α∀problem,␈α∀perhaps␈α∀one␈α∀for␈α∀which␈α∀human␈α∀expertise␈α∀is␈α∀in␈α∀short
␈↓ α←␈↓␈↓12    BACKGROUND␈↓ 
#2-2␈↓

␈↓"β␈↓ α←␈↓supply.␈↓
1␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Several␈α∂programs␈α∂of␈α∂this␈α⊂type␈α∂are␈α∂currently␈α∂under␈α⊂development.␈α∂ A
␈↓ α←␈↓program␈α∞for␈α∞diagnosis␈α∞of␈α
glaucoma␈α∞is␈α∞described␈α∞in␈α∞[Kulikowski73];␈α
internal
␈↓ α←␈↓medicine␈αis␈αthe␈α
domain␈αfor␈αanother␈αeffort␈α
described␈αby␈α[Pople75];␈α
and␈αwork
␈↓ α←␈↓on␈α
a␈αconsultation␈α
program␈α
for␈αelectro-mechanical␈α
repair␈αtasks␈α
is␈α
reported␈αin
␈↓ α←␈↓[Hart75].␈α↔ These␈α↔programs␈α↔all␈α↔rely␈α↔on␈α↔large␈α↔stores␈α↔of␈α↔domain-specific
␈↓ α←␈↓knowledge␈α∂for␈α∂their␈α∂performance,␈α∞and␈α∂thus␈α∂could␈α∂be␈α∂considered␈α∞candidates
␈↓ α←␈↓for␈α∂the␈α⊂kind␈α∂of␈α⊂performance␈α∂program␈α⊂␈↓¬TEIRESIAS␈↓␈α∂has␈α⊂been␈α∂designed␈α⊂to␈α∂help
␈↓ α←␈↓construct.

␈↓"β␈↓ α←␈↓␈↓α2-2-1    Mycin␈↓
␈↓"β␈↓ α←␈↓␈↓ β?One␈α⊗such␈α⊗program,␈α⊗the␈α⊗␈↓¬MYCIN␈↓␈α⊗system␈α↔([Shortliffe76],␈α⊗[Davis77b]),
␈↓ α←␈↓provided␈α
the␈α
context␈αin␈α
which␈α
␈↓¬TEIRESIAS␈↓␈α
was␈αdeveloped␈α
and␈α
played␈α
the␈αrole␈α
of
␈↓ α←␈↓the␈α⊗performance␈α⊗program␈α⊗in␈α⊗Fig.␈α⊗1-3.␈α⊗ ␈↓¬MYCIN␈↓␈α⊗was␈α⊗designed␈α⊗to␈α⊗provide
␈↓ α←␈↓consultative␈α∂advice␈α∂on␈α∂diagnosis␈α∂and␈α∂therapy␈α∂for␈α∂infectious␈α∂diseases.␈α∞ Such
␈↓ α←␈↓advice␈αis␈αoften␈αrequired␈αin␈αthe␈αhospital␈αbecause␈αthe␈αattending␈αphysician␈αmay
␈↓ α←␈↓not␈α∞be␈α∞an␈α∞expert␈α∞on␈α∞infectious␈α∞disease--as,␈α∞for␈α∞example,␈α∞when␈α∞a␈α
cardiology
␈↓ α←␈↓patient␈α↔develops␈α↔an␈α_infection␈α↔after␈α↔heart␈α↔surgery.␈α_Time␈α↔considerations
␈↓ α←␈↓compound␈α∂the␈α∂problem.␈α∞A␈α∂specimen␈α∂(blood,␈α∂urine,␈α∞etc.)␈α∂from␈α∂a␈α∂patient␈α∞can
␈↓ α←␈↓show␈αsome␈α
early␈αevidence␈α
of␈αbacterial␈α
growth␈αwithin␈α
12␈αhours,␈α
but␈α24␈α
to␈α48
␈↓ α←␈↓hours␈α↔(or␈α↔more)␈α↔are␈α↔usually␈α↔required␈α↔for␈α↔positive␈α↔identification.␈α↔ The
␈↓ α←␈↓physician␈αmust␈αtherefore␈αoften␈αdecide␈αin␈αthe␈αabsence␈αof␈αcomplete␈α
information
␈↓ α←␈↓whether␈α⊃or␈α⊃not␈α∩to␈α⊃start␈α⊃treatment␈α∩and␈α⊃what␈α⊃drugs␈α∩to␈α⊃use␈α⊃if␈α∩treatment␈α⊃is
␈↓ α←␈↓required.  Both of these may be difficult questions.
␈↓"β␈↓ α←␈↓␈↓ β?Figs.␈α∪2-1␈α∪and␈α∪2-2␈α∩show␈α∪the␈α∪initial␈α∪and␈α∩final␈α∪parts␈α∪of␈α∪a␈α∩sample
␈↓ α←␈↓interaction␈αbetween␈αa␈αphysician␈αand␈αthe␈αprogram␈α(italicized␈αcomments␈αat␈αthe
␈↓ α←␈↓right␈α
provide␈α
additional␈α∞commentary␈α
but␈α
are␈α
not␈α∞part␈α
of␈α
the␈α∞actual␈α
dialog).
␈↓ α←␈↓␈↓¬MYCIN␈↓␈αin␈α
effect␈α``interviews''␈α
the␈αdoctor␈α
about␈αhis␈α
patient,␈αcollecting␈α
information
␈↓ α←␈↓that will allow it to infer the diagnosis and select an appropriate therapy.





␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[1]␈αThe␈αconcept␈αis␈αdefined␈αhere␈αin␈αterms␈αthat␈αassume␈αthe␈αexistence␈αof␈αhuman
␈↓ α←␈↓experts␈αto␈αinsure␈αthat␈αknowledge␈αin␈αthe␈αdomain␈αis␈αsufficiently␈αadvanced␈αthat
␈↓ α←␈↓it␈αcan␈αreliably␈αsupport␈αhigh␈αperformance.␈α Human␈αexperts␈αalso␈αsupply␈αone␈αof
␈↓ α←␈↓the␈α∞best␈α∞sources␈α∞for␈α∞this␈α∞knowledge␈α∞(learning␈α∞from␈α∞textbooks␈α∞has␈α∂also␈α∞been
␈↓ α←␈↓considered)␈α∞and␈α∞offer␈α∞a␈α∞standard␈α∞by␈α∞which␈α∞to␈α∞gauge␈α∞program␈α
performance.
␈↓ α←␈↓The␈αdefinition␈αincludes␈αonly␈α
cognitive␈αtasks,␈αto␈αminimize␈αcomplicating␈α
factors
␈↓ α←␈↓that␈α⊂result␈α⊂from␈α∂the␈α⊂reliance␈α⊂of␈α∂human␈α⊂performance␈α⊂on␈α⊂highly␈α∂specialized
␈↓ α←␈↓processors (as appears to be true for vision and manipulation).
␈↓ α←␈↓␈↓2-2␈↓ πTCOMPUTER CONSULTANTS    13␈↓

␈↓"β␈↓ α←␈↓	--------PATIENT-1--------
␈↓ α←␈↓	1) Patient's name: (first-last)
␈↓ α←␈↓	**␈↓αFRED SMITH ␈↓	                       ␈↓ λW␈↓↓{Background data.}␈↓	
␈↓ α←␈↓	2) Sex:
␈↓ α←␈↓	**␈↓αMALE ␈↓	
␈↓ α←␈↓	3) Age:
␈↓ α←␈↓	**␈↓α55 ␈↓	
␈↓ α←␈↓	4) Have you been able to obtain positive cultures from a site
␈↓ α←␈↓	   at which Fred Smith has an infection?
␈↓ α←␈↓	**␈↓αYES␈↓	
␈↓ α←␈↓	  --------INFECTION-1--------
␈↓ α←␈↓	  5) What is the infection?
␈↓ α←␈↓	  **␈↓αPRIMARY-BACTEREMIA ␈↓	
␈↓ α←␈↓	  6) Please give date and time when signs or symptoms of the
␈↓ α←␈↓	     primary-bacteremia (INFECTION-1) first appeared.
␈↓ α←␈↓	  **␈↓α MAY 5, 1977  18:00␈↓	
␈↓ α←␈↓	  The most recent positive culture associated with the primary-
␈↓ α←␈↓	  bacteremia (INFECTION-1) will be referred to as:
␈↓ α←␈↓	    --------CULTURE-1--------
␈↓ α←␈↓	    7) From what site was the specimen for CULTURE-1 taken?
␈↓ α←␈↓	    **␈↓αBLOOD ␈↓	
␈↓ α←␈↓	    8) Please give the date and time when this blood culture
␈↓ α←␈↓	       (CULTURE-1) was obtained. (mo/da/yr)
␈↓ α←␈↓	    **␈↓α MAY 9, 1975 ␈↓	
␈↓ α←␈↓	    The first significant organism from this blood culture
␈↓ α←␈↓	    (CULTURE-1) will be referred to as:
␈↓ α←␈↓	      --------ORGANISM-1--------
␈↓ α←␈↓	      9) Enter the identity  of ORGANISM-1:
␈↓ α←␈↓	      **␈↓αUNKNOWN ␈↓	
␈↓ α←␈↓	␈↓ ∧b␈↓↓{Typically the identity is not known and must be inferred.}␈↓	
␈↓ α←␈↓	      10) Is ORGANISM-1 a rod or coccus (etc.):
␈↓ α←␈↓	      **␈↓αROD ␈↓	
␈↓ α←␈↓	      11) The gram stain of ORGANISM-1:
␈↓ α←␈↓	      **␈↓αGRAMNEG ␈↓	
␈↓ α←␈↓	      12) Have there been POSITIVE cultures from sites which
␈↓ α←␈↓	          are NOT associated with infections about which you
␈↓ α←␈↓	          are seeking advice?
␈↓ α←␈↓	      **␈↓αNO␈↓	
␈↓ α←␈↓	      13) Did ORGANISM-1 grow aerobically?
␈↓ α←␈↓	      **␈↓αYES (.8)␈↓	
␈↓ α←␈↓	              ...

␈↓ α←␈↓↓␈↓ βW{More␈α∂questions␈α∂follow,␈α∞to␈α∂gather␈α∂sufficient␈α∂information␈α∞to
␈↓ α←␈↓↓␈↓ βWinfer possible identities of the infecting organisms.}

␈↓ α←␈↓α␈↓ ∧⊗Fig. 2-1.    The initial part of a consultation.    
␈↓ α←␈↓␈↓14    BACKGROUND␈↓ 
#2-2␈↓

␈↓"β␈↓ α←␈↓	My therapy recommendation will be based on the following
␈↓ α←␈↓	probable infection(s) and potential causative organism(s):

␈↓ α←␈↓	INFECTION-1 is PRIMARY-BACTEREMIA       ␈↓ λO␈↓↓{Possible identities.}␈↓	
␈↓ α←␈↓	<1>  The identity of ORGANISM-1 may be PSEUDOMONAS-AERUGINOSA
␈↓ α←␈↓	<2>  The identity of ORGANISM-1 may be KLEBSIELLA-PNEUMONIAE
␈↓ α←␈↓	<3>  The identity of ORGANISM-1 may be E.COLI
␈↓ α←␈↓	<4>  The identity of ORGANISM-1 may be BACTEROIDES-FRAGILIS
␈↓ α←␈↓	<5>  The identity of ORGANISM-1 may be ENTEROBACTER
␈↓ α←␈↓	<6>  The identity of ORGANISM-1 may be PROTEUS-NON-MIRABILIS

␈↓ α←␈↓	46) On a scale of 0 to 4 (higher numbers indicate increasing
␈↓ α←␈↓	    severity), how would you rate Fred Smith's degree of
␈↓ α←␈↓	    sickness?
␈↓ α←␈↓	**␈↓α3␈↓	
␈↓ α←␈↓	47) Does Fred Smith have a clinically significant allergic
␈↓ α←␈↓	    reaction to any antimicrobial agent?
␈↓ α←␈↓	**␈↓αNO ␈↓	
␈↓ α←␈↓	48) Fred Smith's weight in kilograms (or <number> POUNDS):
␈↓ α←␈↓	**␈↓α70␈↓	
␈↓ α←␈↓	                ␈↓ ¬6␈↓↓{Note that 2 drugs cover for 6 potential organisms.}␈↓	
␈↓ α←␈↓	[Rec 1] My preferred therapy recommendation is as follows:
␈↓ α←␈↓	   In order to cover for Items <1 2 3 5 6>:
␈↓ α←␈↓	        Give: GENTAMICIN
␈↓ α←␈↓	        Dose: 119 mg (1.7 mg/kg) q8h IV [or IM] for 10 days
␈↓ α←␈↓	        Comments: Modify dose in renal failure
␈↓ α←␈↓	   In order to cover for Item <4>:
␈↓ α←␈↓	        Give: CLINDAMYCIN
␈↓ α←␈↓	        Dose: 595 mg (8.5 mg/kg) q6h IV [or IM] for 14 days
␈↓ α←␈↓	        Comments: If diarrhea or other GI symptoms develop,
␈↓ α←␈↓	                  patient should be evaluated for possible
␈↓ α←␈↓	                  pseudomembranous colitis.

␈↓ α←␈↓	Do you wish to see the next choice therapy?
␈↓ α←␈↓	**␈↓α YES␈↓	      ␈↓ π∨␈↓↓{Alternative choices are available.}␈↓	
␈↓ α←␈↓	Which drug do you wish eliminated from consideration?
␈↓ α←␈↓	**␈↓αCLINDAMYCIN␈↓	
␈↓ α←␈↓	[Rec 2] Next best therapy recommendation:
␈↓ α←␈↓	   In order to cover for Items <2 3 4 5 6>:
␈↓ α←␈↓	        Give: CHLORAMPHENICOL
␈↓ α←␈↓	        Dose: 1 gm (15 mg/kg) q6h IV for 14 days
␈↓ α←␈↓	        Comments: Monitor patient's white count
␈↓ α←␈↓	   In order to cover for Item <1>:
␈↓ α←␈↓	        Give: GENTAMICIN
␈↓ α←␈↓	        Dose: 119 mg (1.7 mg/kg) q8h IV [or IM] for 10 days
␈↓ α←␈↓	        Comments: Modify dose in renal failure


␈↓ α←␈↓α␈↓ ∧≡Fig. 2-2.    The final part of a consultation.    
␈↓ α←␈↓␈↓2-3␈↓ ε]TEIRESIAS:  SYSTEM ORGANIZATION    15␈↓

␈↓"β␈↓ α←␈↓␈↓α2-3    TEIRESIAS:  SYSTEM ORGANIZATION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈α∞we␈α
have␈α∞noted,␈α
␈↓¬TEIRESIAS␈↓␈α∞provides␈α
a␈α∞number␈α
of␈α∞tools␈α∞designed␈α
to
␈↓ α←␈↓aid␈α
in␈α∞the␈α
construction,␈α
maintenance,␈α∞and␈α
use␈α
of␈α∞the␈α
knowledge␈α
base␈α∞in␈α
the
␈↓ α←␈↓performance␈α
program.␈α
 The␈α
overall␈α
structure␈α
of␈α
the␈α
system␈α
is␈α
shown␈α
in␈α
Fig.
␈↓ α←␈↓2-3,␈α
a␈αmore␈α
detailed␈α
version␈αof␈α
Fig.␈α1-3.␈α
 In␈α
order␈αto␈α
focus␈αon␈α
the␈α
issues␈αof
␈↓ α←␈↓knowledge␈α⊂acquisition,␈α⊃explanation,␈α⊂etc.,␈α⊃covered␈α⊂in␈α⊃later␈α⊂chapters,␈α⊃we␈α⊂will
␈↓ α←␈↓henceforth␈α⊃adopt␈α⊃the␈α⊃perspective␈α⊃suggested␈α⊃in␈α⊃Fig.␈α⊃2-3,␈α⊃and␈α⊃consider␈α⊃the
␈↓ α←␈↓performance␈αprogram␈αas␈αthe␈αsimple␈αentity␈αindicated␈αthere.␈α ␈↓¬MYCIN␈↓␈αis␈α
of␈αcourse
␈↓ α←␈↓more␈α∞complex,␈α∞but␈α∞this␈α∞abstraction␈α∞contains␈α∞all␈α∞the␈α∞detail␈α∞necessary␈α∞for␈α∞our
␈↓ α←␈↓purposes.


␈↓"␈↓ α←␈↓∧         ⊂ααααααααααααααααααααααααααααααααααααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧         ~         TEIRESIAS                                 ~
␈↓"␈↓ α←␈↓∧         ~                            ⊂ααααααααααααααααααα⊃  ~
␈↓"␈↓ α←␈↓∧         ~                            ~PERFORMANCE PROGRAM~  ~
␈↓"␈↓ α←␈↓∧         ~  ⊂ααααααααααααα⊃           ~  ⊂ααααααααααα⊃    ~  ~
␈↓"␈↓ α←␈↓∧  E  ← α α α~ EXPLANATION ~← α α α α α~α ~ INFERENCE ~    ~  ~
␈↓"␈↓ α←␈↓∧         ~  ~             ~           ~  ~  ENGINE   ~    ~  ~
␈↓"␈↓ α←␈↓∧  X      ~  %ααααααααααααα$           ~  %ααααααααααα$    ~  ~
␈↓"␈↓ α←␈↓∧         ~                            ~  ⊂ααααααααααα⊃    ~  ~
␈↓"␈↓ α←␈↓∧  P      ~  ⊂ααααααααααααα⊃           ~  ~ KNOWLEDGE ~    ~  ~
␈↓"␈↓ α←␈↓∧     α α α →~ KNOWLEDGE   ~α α α α α α~→ ~   BASE    ~    ~  ~
␈↓"␈↓ α←␈↓∧  E      ~  ~ ACQUISITION ~           ~  %ααααααααααα$    ~  ~
␈↓"␈↓ α←␈↓∧         ~  %ααααααααααααα$           ~         ↑         ~  ~
␈↓"␈↓ α←␈↓∧  R      ~                            %ααααααααααααααααααα$  ~
␈↓"␈↓ α←␈↓∧     α ⊃ ~                                      ~            ~
␈↓"␈↓ α←␈↓∧  T      ~              ⊂αααααααααααα⊃                       ~
␈↓"␈↓ α←␈↓∧       % α α α α α α α →~ STRATEGIES ~α α α α α $            ~
␈↓"␈↓ α←␈↓∧         ~              %αααααααααααα$                       ~
␈↓"␈↓ α←␈↓∧         %ααααααααααααααααααααααααααααααααααααααααααααααααααα$


␈↓"␈↓ α←␈↓α␈↓ ∧$Fig. 2-3.    Teiresias:  System organization.    

␈↓"β␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓'s␈α≥␈↓↓explanation␈↓␈α≥facility␈α≡(see␈α≥chapter␈α≥3)␈α≥uses␈α≡both␈α≥the
␈↓ α←␈↓information␈α∞in␈α∞the␈α∂knowledge␈α∞base␈α∞and␈α∞an␈α∂understanding␈α∞of␈α∞the␈α∂design␈α∞of
␈↓ α←␈↓the␈α_inference␈α↔engine␈α_to␈α_provide␈α↔the␈α_expert␈α↔with␈α_explanations␈α_of␈α↔the
␈↓ α←␈↓performance␈α∨program's␈α≡behavior.␈α∨ The␈α≡␈↓↓knowledge␈α∨acquisition␈↓␈α≡facility
␈↓ α←␈↓(chapters␈α∞4-6)␈α∞allows␈α∞the␈α∞expert␈α∞to␈α∞transfer␈α∞his␈α∞knowledge␈α∞of␈α∞the␈α∂field␈α∞into
␈↓ α←␈↓the␈α→knowledge␈α→base,␈α_in␈α→order␈α→to␈α_increase␈α→the␈α→performance␈α_program's
␈↓ α←␈↓competence.␈α∂ Finally,␈α∞the␈α∂base␈α∞of␈α∂␈↓↓strategy␈α∞knowledge␈↓␈α∂(chapter␈α∞7)␈α∂provides␈α∞a
␈↓ α←␈↓mechanism␈αfor␈αexpressing␈αstrategies␈αconcerning␈αthe␈αuse␈αof␈αinformation␈αin␈αthe
␈↓ α←␈↓knowledge base.␈↓
2␈↓
␈↓"β␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓␈α∀is␈α∀written␈α∀in␈α∃␈↓¬INTERLISP␈↓,␈α∀an␈α∀advanced␈α∀dialect␈α∀of␈α∃the␈α∀␈↓¬LISP␈↓
␈↓ α←␈↓language,␈α∀and␈α∀runs␈α∀on␈α∀a␈α∪DEC␈α∀PDP-10␈α∀under␈α∀Tenex.␈α∀ The␈α∪knowledge

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[2]␈αIdeally,␈αit␈αshould␈αbe␈αpossible␈αfor␈αthe␈αexpert␈αto␈αadd␈αstrategy␈α
knowledge␈αto
␈↓ α←␈↓the␈α∞system␈α∂using␈α∞the␈α∞knowledge␈α∂acquisition␈α∞facility,␈α∞but␈α∂this␈α∞has␈α∞not␈α∂as␈α∞yet
␈↓ α←␈↓been implemented.  See chapter 7 for details.
␈↓ α←␈↓␈↓16    BACKGROUND␈↓ 
#2-3␈↓

␈↓"β␈↓ α←␈↓acquisition␈α_program␈α_occupies␈α_approximately␈α_40,000␈α_(36-bit)␈α_words;␈α↔the
␈↓ α←␈↓explanation program, 10,000; and the strategy knowledge base, 5,000.

␈↓"β␈↓ α←␈↓␈↓α2-3-1    Performance program:  Knowledge base organization␈↓
␈↓"β␈↓ α←␈↓␈↓ εQIn time you will know all with certainty. 
␈↓"β␈↓ α←␈↓␈↓ 	gline 613
␈↓"β␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓␈α
is␈α
designed␈α
to␈α
deal␈α∞with␈α
knowledge␈α
encoded␈α
in␈α
the␈α∞form␈α
of
␈↓ α←␈↓inference␈α⊗rules␈α⊗of␈α⊗the␈α⊗sort␈α⊗shown␈α⊗in␈α⊗Fig.␈α⊗2-4.␈α⊗ The␈α⊗rules␈α↔are␈α⊗stored
␈↓ α←␈↓internally␈α∂in␈α⊂the␈α∂␈↓¬INTERLISP␈↓␈α⊂form␈α∂shown,␈α⊂from␈α∂which␈α⊂the␈α∂English␈α⊂version␈α∂is
␈↓ α←␈↓generated.␈α∩ Each␈α∩rule␈α⊃is␈α∩a␈α∩single␈α⊃``chunk''␈α∩of␈α∩domain-specific␈α⊃information
␈↓ α←␈↓indicating␈α∀an␈α∃␈↓↓action␈↓␈α∀(in␈α∃this␈α∀case␈α∀a␈α∃conclusion)␈α∀that␈α∃is␈α∀justified␈α∃if␈α∀the
␈↓ α←␈↓conditions specified in the ␈↓↓premise␈↓ are fulfilled.

␈↓"β␈↓ α←␈↓	RULE050
␈↓"β␈↓ α←␈↓	-------

␈↓"β␈↓ α←␈↓	If  1) the infection is primary-bacteremia, and
␈↓"β␈↓ α←␈↓	    2) the site of the culture is one of the sterile sites, and
␈↓"β␈↓ α←␈↓	    3) the suspected portal of entry of the organism is the
␈↓"β␈↓ α←␈↓	       gastrointestinal tract,
␈↓"β␈↓ α←␈↓	then there is suggestive evidence (.7) that the identity of the
␈↓"β␈↓ α←␈↓	     organism is bacteroides.


␈↓"β␈↓ α←␈↓	PREMISE    ($AND (SAME CNTXT INFECT PRIMARY-BACTEREMIA)
␈↓"β␈↓ α←␈↓	                 (MEMBF CNTXT SITE STERILESITES)
␈↓"β␈↓ α←␈↓	                 (SAME CNTXT PORTAL GI))
␈↓"β␈↓ α←␈↓	ACTION     (CONCLUDE CNTXT IDENT BACTEROIDES TALLY .7)

␈↓"β␈↓ α←␈↓α␈↓ ∧␈Fig. 2-4.    An inference rule.    

␈↓ α←␈↓Note␈αthat␈αthe␈αrules␈αare␈αjudgmental,␈αthat␈αis,␈αthey␈αmake␈αinexact␈αinferences.␈α In
␈↓ α←␈↓the␈α
case␈α
of␈αthe␈α
rule␈α
in␈α
Fig.␈α2-4,␈α
for␈α
instance,␈α
the␈αevidence␈α
cited␈α
in␈αthe␈α
premise
␈↓ α←␈↓is␈α
enough␈α
to␈α
assert␈α
the␈α
conclusion␈α
shown␈α
with␈α
a␈α
mild␈α
degree␈α
of␈αconfidence: ␈α
.7
␈↓ α←␈↓out␈αof␈α1.0.␈α This␈αnumber␈αis␈αcalled␈α
a␈α``certainty␈αfactor,''␈αor␈αCF,␈αand␈αembodies␈α
a
␈↓ α←␈↓model␈α
of␈α
confirmation␈α
described␈α
in␈α
[Shortliffe75b].␈α
 The␈α
details␈α
of␈αthis␈α
model
␈↓ α←␈↓need␈α⊃not␈α⊃concern␈α⊃us␈α∩here;␈α⊃we␈α⊃need␈α⊃only␈α⊃note␈α∩that␈α⊃rules␈α⊃in␈α⊃this␈α∩case␈α⊃are
␈↓ α←␈↓typically inexact inferences, rather than statements made with certainty.␈↓
3␈↓

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[3]␈α∪$AND␈α∪(the␈α∪multivalued␈α∪analogue␈α∪of␈α∪the␈α∪Boolean␈α∪AND)␈α∪performs␈α∩a
␈↓ α←␈↓minimization␈α∀operation;␈α∀$OR␈α∀(similar)␈α∀does␈α∀a␈α∀maximization.␈α∃ Note␈α∀that,
␈↓ α←␈↓unlike␈α!standard␈α!probability␈α!theory,␈α $AND␈α!does␈α!not␈α!involve␈α any
␈↓ α←␈↓multiplication␈α
over␈α
its␈α
arguments.␈α
 Since␈α
CFs␈α
are␈α
not␈α
probabilities,␈α
there␈αis␈α
no
␈↓ α←␈↓a priori␈α⊂reason␈α⊃why␈α⊂a␈α⊃product␈α⊂should␈α⊃be␈α⊂a␈α⊃reasonable␈α⊂number.␈α⊃There␈α⊂is,
␈↓ α←␈↓moreover,␈αa␈αlong-standing␈αconvention␈α
in␈αwork␈αwith␈αmultivalued␈αlogics␈α
which
␈↓ α←␈↓interprets␈α∃AND␈α∃as␈α∃␈↓↓min␈↓␈α∃and␈α∀OR␈α∃as␈α∃␈↓↓max␈↓␈α∃[Lukasciewicz70].␈α∃It␈α∃is␈α∀based
␈↓ α←␈↓primarily␈αon␈αintuitive␈αgrounds: ␈αIf␈αa␈αconclusion␈αrequires␈αall␈αof␈αits␈αantecedents
␈↓ α←␈↓␈↓2-3␈↓ ε]TEIRESIAS:  SYSTEM ORGANIZATION    17␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α∞premise␈α∞of␈α
each␈α∞rule␈α∞is␈α
a␈α∞Boolean␈α∞combination␈α
of␈α∞one␈α∞or␈α
more
␈↓ α←␈↓␈↓↓clauses␈↓,␈α∩each␈α∪of␈α∩which␈α∩is␈α∪constructed␈α∩from␈α∩a␈α∪␈↓↓predicate␈α∩function␈↓␈α∪with␈α∩an
␈↓ α←␈↓␈↓↓associative␈αtriple␈↓␈α
(␈↓↓attribute,␈αobject,␈α
value␈↓)␈αas␈α
its␈αargument.␈α
 Thus␈αeach␈αclause␈α
of
␈↓ α←␈↓a premise has the following four components:

␈↓"β␈↓ α←␈↓∧␈↓ β∂<predicate function>    <object>    <attribute>   <value>

␈↓ α←␈↓For␈α∞the␈α∞third␈α∞clause␈α∂in␈α∞the␈α∞premise␈α∞of␈α∂Fig.␈α∞2-4,␈α∞for␈α∞example,␈α∂the␈α∞predicate
␈↓ α←␈↓function␈α
is␈α␈↓	SAME␈↓,␈α
and␈αthe␈α
triple␈αis␈α
``␈↓↓portal-of-entry␈↓␈αof␈α
␈↓↓organism␈↓␈α
is␈α␈↓↓GI-tract␈↓.'' 
␈↓ α←␈↓(␈↓	CNTXT␈↓␈αis␈αa␈αfree␈αvariable␈αwhich␈αis␈αbound␈αto␈αthe␈αspecific␈αobject␈α[also␈αcalled␈αa
␈↓ α←␈↓``context'']␈α⊂for␈α⊂which␈α⊂the␈α⊂rule␈α⊂is␈α⊂invoked.) ␈α⊂There␈α⊂is␈α⊂a␈α⊂standardized␈α⊂set␈α∂of
␈↓ α←␈↓some␈α~24␈α→domain-independent␈α~predicate␈α→functions␈α~(e.g.,␈α~␈↓	SAME,␈α→KNOWN,
␈↓ α←␈↓	DEFINITE␈↓)␈αand␈αa␈αrange␈αof␈αdomain-specific␈αattributes␈α(e.g.,␈α␈↓	IDENTITY,␈αSITE␈↓),
␈↓ α←␈↓objects␈α∃(e.g.,␈α∃␈↓	ORGANISM,␈α⊗CULTURE␈↓),␈α∃and␈α∃associated␈α∃values␈α⊗(e.g.,␈α∃␈↓	E.COLI,
␈↓ α←␈↓	BLOOD␈↓).␈α∂ These␈α⊂form␈α∂a␈α∂``vocabulary''␈α⊂of␈α∂conceptual␈α∂primitives␈α⊂available␈α∂for
␈↓ α←␈↓use in constructing rules.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∂rule␈α∞premise␈α∂is␈α∞always␈α∂a␈α∞conjunction␈α∂of␈α∞clauses,␈α∂but␈α∂may␈α∞contain
␈↓ α←␈↓arbitrarily␈α∂complex␈α⊂conjunctions␈α∂or␈α∂disjunctions␈α⊂nested␈α∂within␈α⊂each␈α∂clause.
␈↓ α←␈↓(Instead␈αof␈α
writing␈αrules␈α
whose␈αpremise␈αwould␈α
be␈αa␈α
disjunction␈αof␈α
clauses,␈αa
␈↓ α←␈↓separate␈α⊂rule␈α⊂is␈α∂written␈α⊂for␈α⊂each␈α∂clause.)␈α⊂The␈α⊂action␈α∂part␈α⊂indicates␈α⊂one␈α∂or
␈↓ α←␈↓more␈αconclusions␈αthat␈αcan␈αbe␈αdrawn␈αif␈αthe␈αpremises␈αare␈αsatisfied,␈αmaking␈αthe
␈↓ α←␈↓rules purely inferential.
␈↓"β␈↓ α←␈↓␈↓ β?Each␈α∩rule␈α∩is␈α∩intended␈α∪to␈α∩embody␈α∩a␈α∩single,␈α∩independent␈α∪chunk␈α∩of
␈↓ α←␈↓knowledge␈α∪and␈α∩states␈α∪all␈α∩necessary␈α∪information␈α∩explicitly␈α∪in␈α∪the␈α∩premise.
␈↓ α←␈↓Since␈αthe␈α
rule␈αuses␈αa␈α
vocabulary␈αof␈α
concepts␈αcommon␈αto␈α
the␈αdomain,␈αit␈α
forms,
␈↓ α←␈↓by␈α
itself,␈αa␈α
comprehensible␈α
statement␈αof␈α
some␈α
piece␈αof␈α
domain␈αknowledge.␈α
 As
␈↓ α←␈↓will become clear, this characteristic is useful in many ways.
␈↓"β␈↓ α←␈↓␈↓ β?Each␈αrule␈αis␈αhighly␈αstylized,␈αwith␈αthe␈αif/then␈αformat␈αand␈αthe␈αspecified
␈↓ α←␈↓set␈α∞of␈α
available␈α∞primitives.␈α
 While␈α∞the␈α
␈↓¬LISP␈↓␈α∞form␈α
of␈α∞each␈α
is␈α∞executable␈α
code
␈↓ α←␈↓(the␈α
premise,␈α
in␈αfact,␈α
is␈α
simply␈α
EVALuated␈αby␈α
␈↓¬LISP␈↓␈α
to␈αtest␈α
its␈α
truth;␈α
and␈αthe
␈↓ α←␈↓action␈α∪EVALuated␈α∪to␈α∪make␈α∪its␈α∪conclusions),␈α∪this␈α∪tightly␈α∪structured␈α∪form
␈↓ α←␈↓makes␈αit␈αpossible␈αto␈αexamine␈αthe␈αrules␈αas␈αwell␈αas␈αexecute␈αthem.␈α This␈αin␈αturn
␈↓ α←␈↓leads␈α∞to␈α
some␈α∞important␈α∞capabilities,␈α
described␈α∞below.␈α
 As␈α∞one␈α∞example,␈α
the
␈↓ α←␈↓internal␈α
form␈α
(i.e.,␈α␈↓¬LISP␈↓)␈α
can␈α
be␈αtranslated␈α
into␈α
the␈αEnglish␈α
form␈α
shown␈αin␈α
Fig.
␈↓ α←␈↓2-4.
␈↓"β␈↓ α←␈↓␈↓ β?Facts␈α⊂about␈α⊂the␈α⊂world␈α⊂(Fig.␈α⊂2-5)␈α⊂are␈α⊂represented␈α⊂as␈α⊃4-tuples␈α⊂made
␈↓ α←␈↓up␈α∩of␈α∩an␈α∩associative␈α∩triple␈α∩and␈α∩its␈α∩current␈α∩CF.␈α∩ Positive␈α∩CFs␈α∩indicate␈α∩a
␈↓ α←␈↓predominance␈α⊃of␈α⊂evidence␈α⊃confirming␈α⊃a␈α⊂hypothesis;␈α⊃negative␈α⊃CFs␈α⊂indicate
␈↓ α←␈↓predominance of disconfirming evidence.


␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓to␈α∂be␈α∂true,␈α∂then␈α∞it␈α∂is␈α∂a␈α∂relatively␈α∂safe␈α∞and␈α∂conservative␈α∂strategy␈α∂to␈α∂use␈α∞the
␈↓ α←␈↓smallest␈α∞of␈α∞the␈α∞antecedent␈α∞values␈α∞as␈α∞the␈α∞value␈α∞of␈α∞the␈α∞premise.␈α∞ Similarly,␈α∞if
␈↓ α←␈↓any␈α
one␈α
of␈αthe␈α
antecedent␈α
clauses␈αjustifies␈α
the␈α
conclusion,␈α
it␈αis␈α
safe␈α
to␈αtake␈α
the
␈↓ α←␈↓maximum value.
␈↓ α←␈↓␈↓18    BACKGROUND␈↓ 
#2-3␈↓


␈↓"β␈↓ α←␈↓	                (IDENT ORGANISM-2 KLEBSIELLA .25)
␈↓"β␈↓ α←␈↓	                (IDENT ORGANISM-2 E.COLI .83)
␈↓"β␈↓ α←␈↓	                (SENSITIVS ORGANISM-1 PENICILLIN -1.0)

␈↓"β␈↓ α←␈↓α␈↓ ∧αFig. 2-5.    Examples of representation of facts.    

␈↓ α←␈↓Note␈α∪that␈α∪the␈α∪model␈α∪of␈α∀inexact␈α∪logic␈α∪permits␈α∪the␈α∪coexistence␈α∀of␈α∪several
␈↓ α←␈↓plausible␈α∞values␈α∞for␈α∞a␈α∞single␈α∂attribute,␈α∞if␈α∞suggested␈α∞by␈α∞the␈α∂evidence.␈α∞ Thus,
␈↓ α←␈↓for␈αexample,␈αafter␈αattempting␈αto␈αdeduce␈αthe␈αidentity␈α(␈↓	IDENT␈↓)␈αof␈αan␈αorganism,
␈↓ α←␈↓␈↓¬MYCIN␈↓␈α∂may␈α∂have␈α∂concluded␈α∂(correctly)␈α∂that␈α∂there␈α∂is␈α∂evidence␈α∂both␈α∂for␈α∂E.coli
␈↓ α←␈↓and for Klebsiella.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α
are␈α
thus␈α∞two␈α
major␈α
forms␈α∞of␈α
knowledge␈α
representation␈α∞in␈α
use
␈↓ α←␈↓in␈α⊂the␈α⊂performance␈α∂program:␈α⊂(␈↓↓i␈↓)␈α⊂the␈α∂attributes,␈α⊂objects,␈α⊂and␈α∂values--which
␈↓ α←␈↓form␈α∩a␈α∩vocabulary␈α∩of␈α∩domain-specific␈α∩conceptual␈α∩primitives,␈α∩and␈α∩(␈↓↓ii␈↓)␈α∩the
␈↓ α←␈↓inference␈α≤rules␈α≤expressed␈α≤in␈α≠terms␈α≤of␈α≤these␈α≤primitives.␈α≤ There␈α≠are,
␈↓ α←␈↓correspondingly,␈α∞two␈α∂forms␈α∞of␈α∂knowledge␈α∞acquisition: ␈α∂(␈↓↓i␈↓)␈α∞the␈α∂acquisition␈α∞of
␈↓ α←␈↓new␈α~primitives--to␈α~expand␈α~the␈α~performance␈α~program's␈α≠vocabulary␈α~of
␈↓ α←␈↓concepts,␈αand␈α
(␈↓↓ii␈↓)␈αthe␈αacquisition␈α
of␈αnew␈αrules␈α
expressed␈αin␈αterms␈α
of␈αexisting
␈↓ α←␈↓primitives.␈α
 ␈↓¬TEIRESIAS␈↓␈α∞makes␈α
possible␈α
both␈α∞of␈α
these;␈α
chapter␈α∞6␈α
deals␈α∞with␈α
the
␈↓ α←␈↓first, and chapter 5 describes the second.

␈↓"β␈↓ α←␈↓␈↓α2-3-2    Performance program:  The inference engine␈↓
␈↓"β␈↓ α←␈↓␈↓ ¬¬Know that I have gone many ways wandering in thought. 
␈↓"β␈↓ α←␈↓␈↓ 	?lines 66-67
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂rules␈α∂are␈α∞invoked␈α∂in␈α∂a␈α∞simple␈α∂backward-chaining␈α∂fashion␈α∞that
␈↓ α←␈↓produces␈α∞an␈α∞exhaustive␈α∞depth-first␈α∂search␈α∞of␈α∞an␈α∞and/or␈α∞goal␈α∂tree.␈α∞ Assume
␈↓ α←␈↓that␈α⊃the␈α∩program␈α⊃is␈α∩attempting␈α⊃to␈α∩determine␈α⊃the␈α∩identity␈α⊃of␈α∩an␈α⊃infecting
␈↓ α←␈↓organism.␈α
 It␈α∞retrieves␈α
all␈α
the␈α∞rules␈α
that␈α
make␈α∞a␈α
conclusion␈α
about␈α∞that␈α
topic
␈↓ α←␈↓(i.e.,␈α⊃they␈α⊃mention␈α⊃␈↓	IDENT␈↓␈α⊃in␈α⊃their␈α⊃action),␈α⊃and␈α⊃invokes␈α⊃each␈α⊃one␈α∩in␈α⊃turn,
␈↓ α←␈↓evaluating␈α
each␈α
premise␈α
to␈α
see␈αif␈α
the␈α
conditions␈α
specified␈α
have␈α
been␈αmet.␈α
 For
␈↓ α←␈↓the␈α∞rule␈α
in␈α∞Fig.␈α
2-4,␈α∞this␈α
process␈α∞would␈α
begin␈α∞with␈α
determining␈α∞the␈α∞type␈α
of
␈↓ α←␈↓the infection.  This, in turn, is set up as a subgoal and the process recurs.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩search␈α⊃is␈α∩thus␈α⊃depth-first␈α∩(because␈α⊃each␈α∩premise␈α∩condition␈α⊃is
␈↓ α←␈↓thoroughly␈αexplored␈αin␈αturn);␈αthe␈αtree␈αthat␈αis␈αsprouted␈αis␈αan␈αand/or␈αgoal␈αtree
␈↓ α←␈↓(because␈α
rules␈α
may␈α
have␈α
OR␈α
conditions␈αin␈α
their␈α
premise);␈α
and␈α
the␈α
search␈αis
␈↓ α←␈↓exhaustive␈α(because␈αthe␈αrules␈αare␈αinexact;␈αso␈αthat␈αeven␈αif␈αone␈αsucceeds,␈αit␈αwas
␈↓ α←␈↓deemed␈αto␈αbe␈αa␈αwisely␈αconservative␈αstrategy␈αto␈αcontinue␈αto␈αcollect␈αall␈αevidence
␈↓ α←␈↓about the subgoal.)
␈↓"β␈↓ α←␈↓␈↓ β?Note␈α⊃that␈α⊂the␈α⊃subgoal␈α⊃that␈α⊂is␈α⊃set␈α⊂up␈α⊃is␈α⊃a␈α⊂generalized␈α⊃form␈α⊃of␈α⊂the
␈↓ α←␈↓original␈αgoal.␈αThus,␈αfor␈αthe␈αfirst␈αclause␈αin␈αFig.␈α2-4␈α(``the␈αinfection␈αis␈α
primary-
␈↓ α←␈↓bacteremia''),␈α∂the␈α∂subgoal␈α∂set␈α∂up␈α⊂is␈α∂``determine␈α∂the␈α∂type␈α∂of␈α⊂infection.'' ␈α∂The
␈↓ α←␈↓subgoal␈αis␈αtherefore␈α
always␈αof␈αthe␈αform␈α
``determine␈αthe␈αvalue␈α
of␈α<attribute>''
␈↓ α←␈↓rather␈α⊃than␈α⊃``determine␈α⊂whether␈α⊃the␈α⊃<attribute>␈α⊂is␈α⊃equal␈α⊃to␈α⊃<value>.'' ␈α⊂By
␈↓ α←␈↓setting␈α
up␈α
the␈α∞generalized␈α
goal␈α
of␈α
collecting␈α∞all␈α
evidence␈α
about␈α∞an␈α
attribute,
␈↓ α←␈↓the␈α≠performance␈α≠program␈α≠effectively␈α≠exhausts␈α≠each␈α≠subject␈α≠as␈α≤it␈α≠is
␈↓ α←␈↓␈↓2-3␈↓ ε]TEIRESIAS:  SYSTEM ORGANIZATION    19␈↓

␈↓"β␈↓ α←␈↓encountered,␈α∂and␈α∂thus␈α∂tends␈α∂to␈α∞group␈α∂together␈α∂all␈α∂questions␈α∂about␈α∂a␈α∞given
␈↓ α←␈↓topic.␈α
 This␈α
results␈α
in␈α
a␈α
system␈α
that␈α
displays␈α
a␈α
much␈α
more␈α
focused,␈α
methodical
␈↓ α←␈↓approach␈α~to␈α~the␈α~task,␈α~which␈α~is␈α~a␈α~distinct␈α~advantage␈α~where␈α→human
␈↓ α←␈↓engineering␈αconsiderations␈αare␈αimportant.␈α The␈αcost␈αis␈αthe␈αeffort␈αof␈αdeducing
␈↓ α←␈↓or␈α∂collecting␈α∂information␈α∂that␈α∂is␈α∂not␈α∂strictly␈α∂necessary.␈α∂ However,␈α∂since␈α∞this
␈↓ α←␈↓occurs␈αrarely--only␈αwhen␈αthe␈α<attribute>␈αcan␈αbe␈αdeduced␈αwith␈αcertainty␈αto␈αbe
␈↓ α←␈↓the␈α
<value>␈αnamed␈α
in␈α
the␈αoriginal␈α
goal--it␈αhas␈α
not␈α
proven␈αto␈α
be␈α
a␈αproblem
␈↓ α←␈↓in practice.
␈↓"β␈↓ α←␈↓␈↓ β?If␈αafter␈αtrying␈αall␈αrelevant␈αrules␈α(referred␈αto␈αas␈α``tracing''␈αthe␈αsubgoal),
␈↓ α←␈↓the␈α
total␈α
weight␈α
of␈αthe␈α
evidence␈α
about␈α
a␈αhypothesis␈α
falls␈α
between␈α
-.2␈α
and␈α.2
␈↓ α←␈↓(an␈αempirical␈αthreshold),␈αthe␈αanswer␈αis␈αregarded␈αas␈αstill␈αunknown.␈α This␈αmay
␈↓ α←␈↓happen␈αif␈αno␈αrules␈αare␈αapplicable,␈αif␈αthe␈αapplicable␈αrules␈αare␈αtoo␈αweak,␈αif␈αthe
␈↓ α←␈↓effects␈α⊂of␈α⊂several␈α⊂rules␈α∂offset␈α⊂each␈α⊂other,␈α⊂or␈α⊂if␈α∂there␈α⊂are␈α⊂no␈α⊂rules␈α⊂for␈α∂this
␈↓ α←␈↓subgoal␈α
at␈αall.␈α
 In␈αany␈α
of␈α
these␈αcases,␈α
when␈αthe␈α
system␈α
is␈αunable␈α
to␈αdeduce␈α
the
␈↓ α←␈↓answer,␈αit␈α
asks␈αthe␈α
user␈αfor␈α
the␈αvalue␈α
of␈αthe␈α
subgoal␈α(using␈α
a␈αphrase␈α
that␈αis
␈↓ α←␈↓stored along with the attribute itself).
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∞strategy␈α∞of␈α
always␈α∞attempting␈α∞to␈α∞deduce␈α
the␈α∞value␈α∞of␈α∞a␈α
subgoal
␈↓ α←␈↓and␈α
asking␈α
the␈α
user␈α
only␈α∞when␈α
deduction␈α
fails,␈α
insures␈α
a␈α∞minimum␈α
number
␈↓ α←␈↓of␈α∂questions.␈α∂ It␈α∂would␈α∂also␈α∂mean,␈α∂however,␈α∂that␈α∂work␈α∂might␈α⊂be␈α∂expended
␈↓ α←␈↓searching␈α⊃for␈α⊃a␈α⊃subgoal,␈α⊃arriving␈α⊃perhaps␈α⊃at␈α⊃a␈α⊃less␈α⊃than␈α⊃definite␈α⊂answer,
␈↓ α←␈↓when␈α∂the␈α∞user␈α∂might␈α∞already␈α∂know␈α∞the␈α∂answer␈α∞with␈α∂certainty.␈α∂ To␈α∞prevent
␈↓ α←␈↓this␈α
inefficiency,␈α
some␈α
of␈α
the␈αattributes␈α
have␈α
been␈α
labeled␈α
``laboratory␈αdata,''
␈↓ α←␈↓to␈α∩indicate␈α∩that␈α∩they␈α∩represent␈α∩information␈α∩available␈α∩to␈α∩the␈α∩physician␈α⊃as
␈↓ α←␈↓results␈αof␈αquantitative␈αtests.␈α In␈αthese␈αcases␈αthe␈αdeduce-then-ask␈αprocedure␈αis
␈↓ α←␈↓reversed␈α∞and␈α∞the␈α
system␈α∞will␈α∞attempt␈α
to␈α∞deduce␈α∞the␈α
answer␈α∞only␈α∞if␈α∞the␈α
user
␈↓ α←␈↓cannot␈α⊂supply␈α⊂it.␈α⊂ Given␈α⊂the␈α⊂desire␈α⊂to␈α⊂minimize␈α⊂both␈α⊂tree␈α⊂search␈α⊂and␈α⊂the
␈↓ α←␈↓number␈α∂of␈α∂questions␈α∂asked,␈α∂there␈α∞is␈α∂no␈α∂guaranteed␈α∂optimal␈α∂solution␈α∂to␈α∞the
␈↓ α←␈↓problem␈αof␈αdeciding␈αwhen␈αto␈αask␈αfor␈αinformation␈αand␈αwhen␈αto␈αtry␈αto␈αdeduce
␈↓ α←␈↓it.␈α∩ But␈α⊃the␈α∩distinction␈α⊃described␈α∩has␈α⊃performed␈α∩quite␈α⊃well␈α∩and␈α∩seems␈α⊃to
␈↓ α←␈↓embody a very appropriate criterion.
␈↓"β␈↓ α←␈↓␈↓ β?Two␈α⊗other␈α∃additions␈α⊗to␈α∃straightforward␈α⊗tree␈α∃search␈α⊗increase␈α∃the
␈↓ α←␈↓inference␈α∀engine's␈α∀efficiency.␈α∪ First,␈α∀before␈α∀the␈α∀entire␈α∪list␈α∀of␈α∀rules␈α∀for␈α∪a
␈↓ α←␈↓subgoal␈α∞is␈α∞retrieved,␈α∞the␈α∞program␈α∞attempts␈α∞to␈α∞find␈α∞a␈α∞sequence␈α∞of␈α∞rules␈α∞that
␈↓ α←␈↓would␈α∩establish␈α∩the␈α∩goal␈α∩with␈α∩certainty,␈α∩based␈α∩only␈α∩on␈α∩what␈α∩is␈α⊃currently
␈↓ α←␈↓known.␈α Since␈α
this␈αis␈α
a␈αsearch␈α
for␈αa␈α
sequence␈αof␈α
rules␈αwith␈α
CF=1,␈αthe␈αresult␈α
is
␈↓ α←␈↓termed␈αa␈α␈↓↓unity␈αpath␈↓.␈α Besides␈α
efficiency␈αconsiderations,␈αthis␈αprocess␈αoffers␈α
the
␈↓ α←␈↓advantage␈αof␈αallowing␈αthe␈αprogram␈αto␈αmake␈α``common␈αsense''␈αdeductions␈αwith
␈↓ α←␈↓a␈α⊃minimum␈α⊃of␈α⊃effort␈α∩(rules␈α⊃with␈α⊃CF=1␈α⊃are␈α⊃largely␈α∩definitional).␈α⊃ Because
␈↓ α←␈↓there are few such rules in the system, the search is typically very brief.
␈↓"β␈↓ α←␈↓␈↓ β?Second,␈α⊂the␈α⊃inference␈α⊂engine␈α⊃performs␈α⊂a␈α⊃partial␈α⊂evaluation␈α⊃of␈α⊂rule
␈↓ α←␈↓premises.␈α Since␈αmany␈αattributes␈αare␈αfound␈αin␈αseveral␈αrules,␈αthe␈αvalue␈αof␈αone
␈↓ α←␈↓clause␈α⊂(perhaps␈α⊂the␈α⊂last)␈α⊂in␈α⊂a␈α⊂premise␈α⊂may␈α⊂already␈α⊂have␈α⊃been␈α⊂established
␈↓ α←␈↓while␈α
the␈αrest␈α
are␈α
still␈αunknown.␈α
 If␈αthis␈α
clause␈α
alone␈αwould␈α
make␈αthe␈α
premise
␈↓ α←␈↓false,␈αthere␈αis␈αclearly␈αno␈αreason␈αto␈αdo␈αall␈αthe␈αsearch␈αnecessary␈αto␈αestablish␈αthe
␈↓ α←␈↓others.␈α⊂Each␈α∂premise␈α⊂is␈α∂thus␈α⊂``previewed''␈α∂by␈α⊂evaluating␈α∂it␈α⊂on␈α∂the␈α⊂basis␈α∂of
␈↓ α←␈↓␈↓20    BACKGROUND␈↓ 
#2-3␈↓

␈↓"β␈↓ α←␈↓currently␈α⊃available␈α⊃information.␈α∩This␈α⊃produces␈α⊃a␈α⊃Boolean␈α∩combination␈α⊃of
␈↓ α←␈↓TRUEs,␈α∪FALSEs,␈α∩and␈α∪UNKNOWNs;␈α∩straightforward␈α∪simplification␈α∩(e.g.,
␈↓ α←␈↓␈↓	F ∧ U ≡ F␈↓)␈αindicates␈αwhether␈αthe␈αrule␈αis␈αguaranteed␈αto␈αfail.␈α This␈αprocedure
␈↓ α←␈↓is examined in more detail in Section 2-4-4.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃final␈α⊂aspect␈α⊃of␈α⊂the␈α⊃control␈α⊂structure␈α⊃is␈α⊂the␈α⊃tree␈α⊂of␈α⊃objects␈α⊂(or
␈↓ α←␈↓contexts)␈α∀that␈α∪is␈α∀constructed␈α∪dynamically␈α∀from␈α∪a␈α∀fixed␈α∪hierarchy␈α∀as␈α∪the
␈↓ α←␈↓consultation␈α⊂proceeds␈α⊂(Fig.␈α⊂2-6).␈α⊃ This␈α⊂tree␈α⊂serves␈α⊂several␈α⊃purposes.␈α⊂ First,
␈↓ α←␈↓bindings␈α
of␈α
free␈αvariables␈α
in␈α
a␈αrule␈α
are␈α
established␈αby␈α
the␈α
context␈α
in␈αwhich
␈↓ α←␈↓the␈α
rule␈αis␈α
invoked,␈αwith␈α
the␈αstandard␈α
access␈αto␈α
contexts␈αthat␈α
are␈αits␈α
ancestors.
␈↓ α←␈↓Second,␈αsince␈αthis␈αtree␈αis␈αused␈αto␈αrepresent␈αthe␈αrelationships␈αof␈αobjects␈αin␈αthe
␈↓ α←␈↓domain,␈α∂it␈α∂helps␈α∂structure␈α∂the␈α∂consultation␈α∂in␈α∂ways␈α∂already␈α∂familiar␈α⊂to␈α∂the
␈↓ α←␈↓user.␈α∪ For␈α∪example,␈α∪in␈α∪the␈α∀medical␈α∪domain,␈α∪a␈α∪patient␈α∪has␈α∪one␈α∀or␈α∪more
␈↓ α←␈↓infections,␈αeach␈α
of␈αwhich␈α
may␈αhave␈αone␈α
or␈αmore␈α
associated␈αcultures,␈α
each␈αof
␈↓ α←␈↓which in turn may have one or more organisms growing in it, and so on.

␈↓"␈↓ α←␈↓∧                            PATIENT-1

␈↓"␈↓ α←␈↓∧                                ~
␈↓"␈↓ α←␈↓∧                                ~
␈↓"␈↓ α←␈↓∧                                ~

␈↓"␈↓ α←␈↓∧                           INFECTION-1

␈↓"␈↓ α←␈↓∧                              /  \
␈↓"␈↓ α←␈↓∧                             /    \
␈↓"␈↓ α←␈↓∧                            /      \
␈↓"␈↓ α←␈↓∧                           /        \

␈↓"␈↓ α←␈↓∧         PREVIOUS-CULTURE-1          CULTURE-1

␈↓"␈↓ α←␈↓∧              ~                      /    \
␈↓"␈↓ α←␈↓∧              ~                     /      \
␈↓"␈↓ α←␈↓∧              ~                    /        \
␈↓"␈↓ α←␈↓∧              ~                   /          \

␈↓"␈↓ α←␈↓∧     PREVIOUS-ORGANISM-1     ORGANISM-1     ORGANISM-2


␈↓"␈↓ α←␈↓α␈↓ ∧RFig. 2-6.    A sample tree of objects.    


␈↓"β␈↓ α←␈↓␈↓α2-3-3    Domain independence and range of application␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αfundamental␈αdesign␈αand␈αimplementation␈αof␈αthe␈αsystem␈αin␈αFig.␈α2-
␈↓ α←␈↓3␈α
does␈α
not␈α
restrict␈α
its␈α
use␈α
to␈α
medical␈α
domains.␈α
 This␈α
is␈α
due␈α
primarily␈α
to␈αthe
␈↓ α←␈↓modularization␈α∩suggested␈α∩in␈α∩that␈α∩figure␈α∩and␈α∩to␈α∩the␈α∩fact␈α∩that␈α∪␈↓¬TEIRESIAS␈↓␈α∩is
␈↓ α←␈↓oriented solely around the concepts of rules, attribute-object-value triples, etc.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αclear␈αdistinction␈α
between␈αthe␈αinference␈α
engine␈αand␈αthe␈α
knowledge
␈↓ α←␈↓base,␈αfor␈αinstance,␈αmakes␈αit␈αpossible␈αto␈αremove␈αone␈αknowledge␈αbase␈α
from␈αthe
␈↓ α←␈↓performance␈α∂program␈α∞and␈α∂replace␈α∂it␈α∞with␈α∂another.␈α∞ It␈α∂has␈α∂proven␈α∞possible,
␈↓"β␈↓ α←␈↓␈↓2-3␈↓ ε]TEIRESIAS:  SYSTEM ORGANIZATION    21␈↓

␈↓"β␈↓ α←␈↓for␈α∞instance,␈α∞to␈α∞build␈α∞separate␈α∂knowledge␈α∞bases␈α∞for␈α∞such␈α∞disparate␈α∂areas␈α∞as
␈↓ α←␈↓auto␈α⊂mechanics␈α⊂and␈α⊂chemotherapy␈α⊂for␈α∂psychiatry.␈α⊂ In␈α⊂the␈α⊂first␈α⊂such␈α∂effort
␈↓ α←␈↓([vanMelle74]),␈α∪a␈α∩small␈α∪part␈α∩of␈α∪an␈α∩auto␈α∪repair␈α∩manual␈α∪was␈α∪rewritten␈α∩as
␈↓ α←␈↓production␈α⊂rules␈α⊃and␈α⊂inserted␈α⊃in␈α⊂place␈α⊂of␈α⊃the␈α⊂bacteremia␈α⊃knowledge␈α⊂base.
␈↓ α←␈↓What␈α∞resulted␈α∞was␈α
a␈α∞very␈α∞simple␈α∞but␈α
fully␈α∞functional␈α∞consultant␈α∞capable␈α
of
␈↓ α←␈↓diagnosing␈α⊗and␈α∃suggesting␈α⊗remedies␈α⊗for␈α∃problems␈α⊗in␈α∃parts␈α⊗of␈α⊗an␈α∃auto
␈↓ α←␈↓electrical␈α
system.␈α
 More␈αrecently,␈α
a␈α
pilot␈αsystem␈α
for␈α
psychiatric␈α
diagnosis␈αand
␈↓ α←␈↓chemotherapy␈α
was␈α
assembled.␈α
 While␈α
this␈α
program␈α
had␈α
only␈α
50␈α
rules,␈α
it␈αtoo
␈↓ α←␈↓was␈α∂fully␈α∞functional␈α∂and␈α∂displayed␈α∞primitive␈α∂but␈α∂encouraging␈α∞performance.
␈↓ α←␈↓In␈α∃both␈α∀systems,␈α∃all␈α∃of␈α∀the␈α∃established␈α∀explanation␈α∃facilities␈α∃worked␈α∀as
␈↓ α←␈↓designed, without modification.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∂are,␈α∞naturally,␈α∂some␈α∂domains␈α∞that␈α∂might␈α∞be␈α∂less␈α∂profitable␈α∞to
␈↓ α←␈↓explore.␈α∂ One␈α∞of␈α∂the␈α∞interesting␈α∂lessons␈α∂of␈α∞the␈α∂auto␈α∞repair␈α∂system␈α∂was␈α∞that
␈↓ α←␈↓domains␈α⊃with␈α⊃little␈α⊃inexactness␈α⊃in␈α⊃the␈α⊃reasoning␈α⊃process--those␈α⊃for␈α⊂which
␈↓ α←␈↓algorithmic␈α≡diagnostic␈α≡routines␈α≡can␈α≡be␈α≡written--are␈α≡not␈α≥particularly
␈↓ α←␈↓appropriate␈α⊂for␈α⊂this␈α⊂methodology.␈α⊃ The␈α⊂precision␈α⊂in␈α⊂these␈α⊃domains␈α⊂means
␈↓ α←␈↓that␈α∞little␈α∂use␈α∞is␈α∞made␈α∂of␈α∞the␈α∂certainty␈α∞factor␈α∞mechanism,␈α∂and␈α∞many␈α∂of␈α∞the
␈↓ α←␈↓more␈α
complicated␈α
(and␈α
computationally␈αexpensive)␈α
features␈α
go␈α
unused.␈α The
␈↓ α←␈↓effect␈α∞would␈α
be␈α∞akin␈α∞to␈α
swatting␈α∞a␈α
fly␈α∞with␈α∞a␈α
large␈α∞doctoral␈α∞thesis--all␈α
that
␈↓ α←␈↓work and weight are unnecessary when something far simpler would do.
␈↓"β␈↓ α←␈↓␈↓ β?Nor␈αis␈αit␈αreasonable␈αto␈αexpect␈αto␈αbe␈αable␈αto␈αwrite␈αinference␈αrules␈αbuilt
␈↓ α←␈↓from␈α∞attribute-object-value␈α∞triples␈α∂for␈α∞an␈α∞arbitrary␈α∞domain.␈α∂ As␈α∞knowledge
␈↓ α←␈↓in␈α
an␈αarea␈α
accumulates,␈αit␈α
becomes␈αprogressively␈α
more␈αformalized.␈α
 There␈αis␈α
a
␈↓ α←␈↓certain␈αstage␈αin␈αthis␈αformalization␈αprocess␈αwhen␈αit␈αis␈αappropriate␈αto␈αuse␈αrules
␈↓ α←␈↓of␈α
the␈α
sort␈α
shown␈α
above.␈α Earlier␈α
than␈α
this␈α
the␈α
knowledge␈α
is␈αtoo␈α
unstructured;
␈↓ α←␈↓later␈α∩on␈α∩it␈α∪may␈α∩(like␈α∩the␈α∪auto␈α∩repair␈α∩system)␈α∪be␈α∩more␈α∩effective␈α∪to␈α∩write
␈↓ α←␈↓straightforward algorithms.
␈↓"β␈↓ α←␈↓␈↓ β?It␈α∞is␈α
also␈α∞possible␈α
that␈α∞the␈α
knowledge␈α∞in␈α
some␈α∞domains␈α∞is␈α
inherently
␈↓ α←␈↓unsuited␈α_to␈α_a␈α_rule-like␈α_representation,␈α_since␈α_rules␈α_become␈α_increasingly
␈↓ α←␈↓awkward␈α
as␈α
the␈α
number␈α
of␈αpremise␈α
clauses␈α
increases.␈α
 Dealing␈α
with␈αa␈α
number
␈↓ α←␈↓of␈α
interacting␈α∞factors␈α
may␈α∞be␈α
difficult␈α∞for␈α
any␈α∞representation,␈α
but␈α∞given␈α
the
␈↓ α←␈↓reliance␈α⊃here␈α⊃on␈α⊃rules␈α⊃as␈α⊃a␈α⊃medium␈α⊃of␈α⊃communication␈α⊃of␈α⊃knowledge,␈α⊃the
␈↓ α←␈↓problem becomes especially significant.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈αthe␈αcurrent␈αperformance␈αprogram␈αmakes␈αextensive␈αuse␈α
of␈αthe
␈↓ α←␈↓attribute-object-value␈α∂associative␈α∂triple.␈α∂ This␈α∂is␈α∂perhaps␈α∂the␈α∂most␈α∂limiting
␈↓ α←␈↓factor,␈αsince␈α
this␈αkind␈αof␈α
representation␈αis␈αeffective␈α
only␈αin␈α
simpler␈αdomains.
␈↓ α←␈↓It␈α⊂is␈α⊂also,␈α⊂however,␈α⊃a␈α⊂far␈α⊂more␈α⊂implementation-dependent␈α⊂factor␈α⊃than␈α⊂the
␈↓ α←␈↓other␈α∪two␈α∀mentioned.␈α∪ While␈α∀it␈α∪would␈α∀be␈α∪difficult,␈α∀the␈α∪triples␈α∀could␈α∪be
␈↓ α←␈↓replaced␈α↔by␈α↔a␈α_more␈α↔powerful␈α↔representation␈α↔scheme␈α_without␈α↔adversely
␈↓ α←␈↓affecting␈α
the␈αfeasibility␈α
of␈α
the␈αtechniques␈α
for␈αknowledge␈α
acquisition␈α
and␈αuse
␈↓ α←␈↓that are described in subsequent chapters.
␈↓ α←␈↓␈↓22    BACKGROUND␈↓ 
#2-4␈↓

␈↓"β␈↓ α←␈↓␈↓α2-4    PRODUCTION RULES␈↓
␈↓"β␈↓ α←␈↓␈↓ β?While␈αmany␈α
of␈αthe␈α
ideas␈αon␈α
which␈α␈↓¬TEIRESIAS␈↓␈α
is␈αbased␈α
are␈αapplicable␈α
to
␈↓ α←␈↓a␈α∂range␈α⊂of␈α∂knowledge␈α⊂representations,␈α∂its␈α⊂design␈α∂and␈α⊂implementation␈α∂have
␈↓ α←␈↓been␈α
strongly␈α
influenced␈α
by␈α
the␈αuse␈α
of␈α
a␈α
rule-based␈α
encoding␈α
of␈αknowledge.
␈↓ α←␈↓This␈α
section␈αexplores␈α
production␈αrules␈α
from␈α
several␈αperspectives,␈α
to␈αprovide␈α
a
␈↓ α←␈↓basis for some of the issues discussed in later chapters.

␈↓"β␈↓ α←␈↓␈↓α2-4-1    Production rules in general␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Production␈α⊃rules␈α⊃have␈α⊃been␈α⊃the␈α⊃subject␈α⊃of␈α⊃much␈α⊃work␈α∩since␈α⊃their
␈↓ α←␈↓introduction␈α⊗by␈α⊗Post␈α⊗([Post43])␈α⊗as␈α⊗a␈α⊗general␈α⊗computational␈α⊗mechanism.
␈↓ α←␈↓Across␈α
the␈α
wide␈α
range␈α
of␈αvariations␈α
and␈α
techniques␈α
explored,␈α
there␈αappears
␈↓ α←␈↓to␈α∞be␈α
a␈α∞set␈α∞of␈α
fundamental␈α∞characteristics␈α∞intrinsic␈α
to␈α∞the␈α∞methodology.␈α
(An
␈↓ α←␈↓extended discussion of these characteristics can be found in [Davis77a].)
␈↓"β␈↓ α←␈↓␈↓ β?The␈αpresent␈αdiscussion␈αwill␈αbe␈αconcerned␈αwith␈αtwo␈αproblems␈αtypically
␈↓ α←␈↓encountered␈α∞in␈α∞using␈α∞production␈α∞rules.␈α∞ They␈α∞are␈α∞described␈α∞briefly␈α∞here,␈α∞so
␈↓ α←␈↓that␈α∃their␈α∃manifestations␈α∃will␈α∃be␈α∃evident␈α∃in␈α∃later␈α∃discussions␈α∃of␈α∀system
␈↓ α←␈↓performance.␈α
 They␈α∞are␈α
also␈α∞described␈α
in␈α∞terms␈α
which␈α∞should␈α
make␈α∞it␈α
clear
␈↓ α←␈↓that␈αthese␈αproblems␈αare␈αinherent␈αin␈αthe␈αproduction␈αsystem␈αmethodology␈αitself.
␈↓ α←␈↓This␈αwill␈αenable␈αthe␈αreader␈αto␈αdistinguish␈αthem␈αfrom␈αshortcomings␈αthat␈αmay
␈↓ α←␈↓have␈α⊃arisen␈α⊃from␈α⊃our␈α⊃approach␈α⊂to␈α⊃various␈α⊃uses␈α⊃of␈α⊃meta-level␈α⊂knowledge.
␈↓ α←␈↓Thus,␈α
while␈α
␈↓¬TEIRESIAS␈↓␈α
displays␈α
certain␈αlimitations,␈α
some␈α
of␈α
these␈α
are␈α
a␈αresult␈α
of
␈↓ α←␈↓the␈α∂approach,␈α∂while␈α∞others␈α∂are␈α∂a␈α∂legacy␈α∞of␈α∂the␈α∂particular␈α∂representation␈α∞of
␈↓ α←␈↓knowledge chosen early in the development of the performance program.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∂problem␈α∂is␈α∞the␈α∂limit␈α∂on␈α∞the␈α∂amount␈α∂of␈α∞knowledge␈α∂that␈α∂can␈α∞be
␈↓ α←␈↓expressed␈α∞conveniently␈α∞in␈α∞a␈α∞single␈α
rule.␈α∞ Actions␈α∞that␈α∞are␈α∞``larger''␈α∞than␈α
this
␈↓ α←␈↓limit␈αare␈α
often␈αachieved␈αby␈α
the␈αcombined␈αeffects␈α
of␈αseveral␈αrules.␈α
 For␈αmany
␈↓ α←␈↓reasons␈α
(see␈α
[Davis77a],␈αsection␈α
5),␈α
this␈αis␈α
difficult␈α
to␈αdo␈α
and␈α
often␈αproduces
␈↓ α←␈↓opaque␈α⊃results.␈α⊂ We␈α⊃will␈α⊃see␈α⊂that␈α⊃this␈α⊃is␈α⊂a␈α⊃significant␈α⊃consideration␈α⊂when
␈↓ α←␈↓judging the utility of production rules as a knowledge representation.
␈↓"β␈↓ α←␈↓␈↓ β?Second,␈α
there␈α
is␈α
what␈α
has␈α
been␈α
labeled␈α
the␈α
implicit␈α∞context␈α
problem.
␈↓ α←␈↓Perhaps␈α∪the␈α∪simplest␈α∪example␈α∪can␈α∪be␈α∪found␈α∪in␈α∪a␈α∩production-rule-based
␈↓ α←␈↓system␈αthat␈αuses␈αa␈αsequential␈αleft-hand␈αside␈α(LHS)␈αscan--that␈αis,␈αit␈αstarts␈αout
␈↓ α←␈↓at␈α∃the␈α∃beginning␈α∃of␈α∃the␈α⊗rule␈α∃set␈α∃and␈α∃searches␈α∃sequentially␈α⊗through␈α∃it,
␈↓ α←␈↓examining␈α∞the␈α∞LHS␈α∞of␈α∞each␈α∞rule␈α∞until␈α∞one␈α∞is␈α∞found␈α∞that␈α∞meets␈α∞the␈α
desired
␈↓ α←␈↓criteria.  Suppose the LHS of the first three rules are:

␈↓"β␈↓ α←␈↓	R1:                     A ∧ B ∧ C  ␈↓∧==@   ###␈↓	
␈↓"β␈↓ α←␈↓	R2:                             D  ␈↓∧==@   ###␈↓	
␈↓"β␈↓ α←␈↓	R3:                         E ∧ F  ␈↓∧==@   ###␈↓	

␈↓ α←␈↓R3␈αwon't␈αbe␈αtested␈αunless␈αD␈αis␈αfalse␈αand␈αeither␈αA,␈αB,␈αor␈αC␈αis␈αfalse.␈α
The␈αrule
␈↓ α←␈↓thus,␈α⊂effectively,␈α⊂has␈α⊂two␈α⊂extra,␈α⊂implicit␈α⊂clauses␈α⊂of␈α⊂the␈α⊂form␈α⊂␈↓	(not␈α⊃D)␈↓␈α⊂and
␈↓ α←␈↓␈↓	((not␈α
A)␈α
∨␈α
(not␈α∞B)␈α
∨␈α
(not␈α
C))␈↓,␈α
simply␈α∞because␈α
of␈α
its␈α
location␈α
in␈α∞the␈α
rule
␈↓ α←␈↓set.␈α Note␈αthat␈αthe␈αlocation␈αis␈αnot␈α␈↓↓necessarily␈↓␈αcritical--the␈αentire␈αthrust␈αof␈αthe
␈↓ α←␈↓rule␈αmay␈αpossibly␈αbe␈αsummed␈αup␈αin␈αthe␈αpresence␈αof␈αjust␈αE␈αand␈αF.␈α But␈αoften
␈↓ α←␈↓␈↓2-4␈↓ λ∨PRODUCTION RULES    23␈↓

␈↓"β␈↓ α←␈↓the location ␈↓↓is␈↓ an important element and the implicit clauses are significant.␈↓
4␈↓
␈↓"β␈↓ α←␈↓␈↓ β?It␈α
is␈αtempting␈α
to␈α
make␈αuse␈α
of␈αimplicit␈α
context␈α
to␈αtake␈α
advantage␈αof␈α
the
␈↓ α←␈↓resulting␈α⊂conciseness--consider␈α⊂how␈α⊂short␈α⊃R3␈α⊂is␈α⊂and␈α⊂how␈α⊂long␈α⊃an␈α⊂explicit
␈↓ α←␈↓statement␈αof␈αthe␈α99th␈αrule␈αmight␈αbe␈αif␈αit␈αdepended␈αon␈αthe␈αfailure␈αof␈αthe␈αfirst
␈↓ α←␈↓98.
␈↓"β␈↓ α←␈↓␈↓ β?While␈αthe␈αpoint␈αhas␈αbeen␈αillustrated␈α
via␈αan␈αordered␈αrule␈αset,␈αthe␈α
same
␈↓ α←␈↓problem␈α∞can␈α
arise␈α∞from␈α
other␈α∞causes.␈α∞␈↓¬MYCIN␈↓'s␈α
rules␈α∞are␈α
not␈α∞ordered,␈α∞but␈α
they
␈↓ α←␈↓are␈α∀classified␈α∀according␈α∃to␈α∀the␈α∀object␈α∀to␈α∃which␈α∀they␈α∀apply␈α∃and␈α∀similar
␈↓ α←␈↓problems arise from this.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αessential␈αpoint␈αhere␈αis␈αtwofold: ␈αFirst,␈α␈↓↓any␈↓␈αpiece␈αof␈αstructuring␈αin
␈↓ α←␈↓the␈α⊂system␈α⊂carries␈α⊂information.␈α⊂ Second,␈α⊂this␈α⊂information␈α⊂can␈α⊂be␈α∂implicitly
␈↓ α←␈↓passed␈α∩on␈α∩to␈α∩the␈α∪rules␈α∩in␈α∩subtle␈α∩ways.␈α∩We␈α∪will␈α∩see␈α∩examples␈α∩of␈α∪this␈α∩in
␈↓ α←␈↓discussing the organization of meta-rules in chapter 7.

␈↓"β␈↓ α←␈↓␈↓α2-4-2    Production rules as a knowledge representation␈↓
␈↓"β␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓'s␈α'first␈α&goal--putting␈α'a␈α&domain␈α'expert␈α'in␈α&direct
␈↓ α←␈↓communication␈α⊗with␈α⊗a␈α↔high␈α⊗performance␈α⊗program--requires␈α↔a␈α⊗common
␈↓ α←␈↓language␈α
of␈α
communication.␈α
 This␈α
is␈α
especially␈α
important␈α
for␈αthe␈α
specification
␈↓ α←␈↓of␈αknowledge␈αthat␈αthe␈αexpert␈αwants␈αto␈αadd␈αto␈αthe␈αsystem.␈αHe␈αmust␈αbe␈αable␈αto
␈↓ α←␈↓express␈αit␈αin␈αa␈αform␈αthat␈αis␈αthe␈αsame␈αas␈α(or␈αtransformable␈αinto)␈αthe␈αone␈αused
␈↓ α←␈↓internally␈α∞by␈α∞the␈α
program.␈α∞ Thus,␈α∞the␈α∞ease␈α
of␈α∞establishing␈α∞the␈α∞link␈α
between
␈↓ α←␈↓expert␈α∩and␈α⊃program␈α∩is␈α⊃strongly␈α∩affected␈α⊃by␈α∩the␈α∩knowledge␈α⊃representation
␈↓ α←␈↓chosen.
␈↓"β␈↓ α←␈↓␈↓ β?While␈α⊂we␈α∂cannot␈α⊂offer␈α∂formal␈α⊂arguments,␈α∂there␈α⊂are␈α⊂several␈α∂reasons
␈↓ α←␈↓why␈α~production␈α~rules␈α~of␈α→the␈α~sort␈α~shown␈α~in␈α→Fig.␈α~2-4␈α~are␈α~a␈α→useful
␈↓ α←␈↓representation.␈α First,␈αthe␈αgeneral␈αtask␈αof␈αdeduction␈αis␈αone␈αthat␈αfits␈αquite␈αwell
␈↓ α←␈↓into␈α
the␈α
situation/action␈α
character␈α
of␈α
production␈α
rules.␈α
 There␈α
is␈αtherefore␈α
less
␈↓ α←␈↓transformation␈α
necessary␈α
between␈α
the␈α∞knowledge␈α
as␈α
expressed␈α
by␈α∞the␈α
expert
␈↓ α←␈↓and its final encoding inside the system.
␈↓"β␈↓ α←␈↓␈↓ β?Next,␈α~the␈α~rules␈α~are,␈α→by␈α~themselves,␈α~comprehensible␈α~chunks␈α→of
␈↓ α←␈↓knowledge,␈α∩since␈α∩they␈α⊃carry␈α∩in␈α∩their␈α∩premise␈α⊃a␈α∩full␈α∩specification␈α∩of␈α⊃their
␈↓ α←␈↓applicability.␈α∞ Their␈α
independence␈α∞also␈α
facilitates␈α∞the␈α
incremental␈α∞growth␈α
of
␈↓ α←␈↓the␈α∪knowledge␈α∪base:  Rules␈α∪can␈α∪be␈α∪added␈α∪one␈α∪by␈α∪one,␈α∪and␈α∪performance
␈↓ α←␈↓improves with each addition.
␈↓"β␈↓ α←␈↓␈↓ β?Rules␈α∞are␈α∞also␈α∞retrieved␈α
and␈α∞invoked␈α∞on␈α∞the␈α
basis␈α∞of␈α∞the␈α∞content␈α
of
␈↓ α←␈↓their␈α∂right-hand␈α⊂side␈α∂(rather␈α⊂than␈α∂on␈α∂the␈α⊂basis␈α∂of␈α⊂their␈α∂name)␈α⊂and␈α∂never
␈↓ α←␈↓reference␈α
one␈α
another.␈α
 As␈α
a␈α
result,␈αadding␈α
(or␈α
changing)␈α
a␈α
rule␈α
is␈α
a␈αfar␈α
easier
␈↓ α←␈↓task␈α∞than␈α∞would␈α
be␈α∞the␈α∞case␈α
if␈α∞there␈α∞were␈α
extensive␈α∞references␈α∞to␈α∞the␈α
rule's
␈↓ α←␈↓name␈α∪in␈α∩many␈α∪places.␈α∪ Consider␈α∩the␈α∪difficulty,␈α∪by␈α∩contrast,␈α∪of␈α∪editing␈α∩a
␈↓ α←␈↓procedure␈αin␈αan␈α␈↓¬ALGOL␈↓-like␈αlanguage␈αand␈αthen␈αhaving␈αto␈αgo␈αback␈αand␈αfind␈α
all
␈↓ α←␈↓the places it is called to see if those calls are still appropriate.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∀rules␈α∃also␈α∀seem␈α∀to␈α∃capture␈α∀a␈α∀``chunk''␈α∃of␈α∀knowledge␈α∃of␈α∀the
␈↓ α←␈↓appropriate␈αsize.␈αA␈αsignificant␈αamount␈αof␈αknowledge␈αof␈αbacteremia␈αdiagnosis

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[4] Indeed, [Waterman70] uses this idea to great advantage.
␈↓ α←␈↓␈↓24    BACKGROUND␈↓ 
#2-4␈↓

␈↓"β␈↓ α←␈↓and␈α⊃therapy␈α⊃has␈α⊃been␈α⊃encoded␈α⊃in␈α⊃rules␈α⊃that␈α⊃have␈α⊃between␈α⊃two␈α∩and␈α⊃five
␈↓ α←␈↓premise␈α∂clauses␈α∂and␈α∂one␈α∂or␈α∂two␈α∞actions.␈α∂ While␈α∂all␈α∂domains␈α∂may␈α∂not␈α∞offer
␈↓ α←␈↓such␈α
convenient␈α
``bite-size''␈α
chunks,␈α
the␈α
success␈α
of␈α
the␈α
systems␈α
in␈α
the␈α
other␈α
two
␈↓ α←␈↓domains that have been explored is encouraging.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α≠control␈α≠structure--backward␈α≠chaining--also␈α≠appears␈α≠to␈α≠be
␈↓ α←␈↓reasonably␈α∃intuitive.␈α⊗ While␈α∃such␈α∃␈↓↓modus␈α⊗ponens␈↓␈α∃reasoning␈α∃is␈α⊗not␈α∃often
␈↓ α←␈↓recognized␈α∞explicitly,␈α
it␈α∞is␈α
common␈α∞cognitive␈α
behavior␈α∞and␈α∞should␈α
therefore
␈↓ α←␈↓not be alien to the expert.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈αthe␈αrules␈α
are,␈αfor␈αthe␈α
most␈αpart,␈αwhat␈α
may␈αbe␈αlabeled␈αa␈α
``single
␈↓ α←␈↓level''␈α∃mechanism.␈α∃ They␈α∀are␈α∃composed␈α∃of␈α∀elements␈α∃that␈α∃are␈α∀conceptual
␈↓ α←␈↓primitives␈αand␈α
require␈αno␈α
further␈αdecomposition␈α
to␈αbe␈α
understood.␈α It␈αis␈α
clear
␈↓ α←␈↓to␈α
the␈α
user,␈α
for␈α
instance,␈αwhat␈α
is␈α
meant␈α
by␈α
``the␈αsite␈α
is␈α
X,''␈α
or␈α
``conclude␈αthat
␈↓ α←␈↓the␈α
identity␈α
is␈α∞Y.'' ␈α
Compare␈α
this␈α∞with␈α
the␈α
difficulty␈α∞that␈α
would␈α
arise␈α∞if␈α
the
␈↓ α←␈↓action of a rule were the invocation of a deep and complex calculation.
␈↓"β␈↓ α←␈↓␈↓ β?As␈αa␈αresult␈αof␈αall␈αof␈α
these␈αfactors,␈αproduction␈αrules␈αof␈αthe␈α
sort␈αshown
␈↓ α←␈↓above have proven to be an effective and useful representation.

␈↓"β␈↓ α←␈↓␈↓α2-4-3    Impact on knowledge organization␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Production␈α→rules␈α→also␈α→contribute␈α→their␈α→own␈α→perspective␈α→on␈α→the
␈↓ α←␈↓fundamental␈α⊂organization␈α⊂of␈α⊂knowledge␈α⊂in␈α⊂a␈α⊂program,␈α⊂a␈α⊂perspective␈α∂quite
␈↓ α←␈↓different␈α~from␈α→that␈α~associated␈α→with␈α~procedure-oriented␈α→representations.
␈↓ α←␈↓Production␈α∪rules␈α∪tend␈α∀to␈α∪de-emphasize␈α∪the␈α∪hierarchical␈α∀control␈α∪structure
␈↓ α←␈↓natural␈αto␈αprocedural␈αlanguages␈αand␈αsubstitute␈αa␈αsingle,␈αuniform␈αcollection␈αof
␈↓ α←␈↓knowledge␈α
``chunks''␈α
(the␈α
rules).␈α Since␈α
each␈α
rule␈α
is␈αretrieved␈α
on␈α
the␈α
basis␈αof
␈↓ α←␈↓its␈αcontents␈α(as␈α
the␈αrule␈αin␈α
Fig.␈α2-4␈αis␈α
retrieved␈αon␈αthe␈α
basis␈αof␈αits␈α
conclusion),
␈↓ α←␈↓no␈αrule␈αis␈αever␈αcalled␈αdirectly,␈αin␈αthe␈αstyle␈αof␈αprocedures.␈α Thus,␈αthe␈αaddition
␈↓ α←␈↓(or␈αdeletion)␈αof␈αa␈αrule␈αdoes␈αnot␈αrequire␈αthe␈αmodification␈αof␈αany␈αother␈αrule␈αto
␈↓ α←␈↓provide␈αfor␈α(or␈αdelete)␈αa␈αcall␈αto␈αit.␈α The␈αresult␈αis␈αa␈αprogram␈αwhose␈αknowledge
␈↓ α←␈↓base␈αis␈αeasily␈αmodified␈αand␈αwhose␈αbehavior␈αis␈αrelatively␈αstable␈αin␈αthe␈αface␈αof
␈↓ α←␈↓those changes.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α⊂stability␈α⊂might␈α⊂be␈α⊂demonstrated␈α⊂by␈α⊂repeatedly␈α⊂removing␈α⊂rules
␈↓ α←␈↓from␈αa␈αproduction-rule-based␈αprogram.␈α Many␈αsuch␈αsystems␈αwill␈α
continue␈αto
␈↓ α←␈↓display␈α∩some␈α∪sort␈α∩of␈α∪``reasonable''␈α∩behavior,␈α∪up␈α∩to␈α∪a␈α∩point.␈α∪ By␈α∩contrast,
␈↓ α←␈↓adding␈α
a␈α
procedure␈α
to␈α
an␈α
␈↓¬ALGOL␈↓-like␈α
program␈α
requires␈α
modification␈α
of␈α
other
␈↓ α←␈↓parts␈α∞of␈α∞the␈α∞code␈α∞to␈α∞insure␈α∞that␈α∞it␈α∞is␈α∞invoked,␈α∞while␈α∞removing␈α∞an␈α
arbitrary
␈↓ α←␈↓procedure from such a program generally cripples it.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈αthat␈α
the␈αissue␈αhere␈α
is␈αmore␈αthan␈α
simply␈αthe␈α``undefined␈α
function''
␈↓ α←␈↓error␈α∩message␈α⊃that␈α∩would␈α∩result␈α⊃from␈α∩a␈α⊃missing␈α∩procedure.␈α∩The␈α⊃problem
␈↓ α←␈↓persists␈α∩even␈α∪if␈α∩the␈α∩compiler␈α∪or␈α∩interpreter␈α∩is␈α∪altered␈α∩to␈α∪treat␈α∩undefined
␈↓ α←␈↓functions␈αas␈α
no-ops.␈αThe␈α
issue␈αis␈α
a␈αmuch␈α
more␈αfundamental␈α
one␈αconcerning
␈↓ α←␈↓organization␈α∨of␈α≡knowledge: ␈α∨Programs␈α≡written␈α∨in␈α≡procedure-oriented
␈↓ α←␈↓languages␈α∞stress␈α∞the␈α∞kind␈α∞of␈α∞explicit␈α∞passing␈α∞of␈α∞control␈α∞from␈α∞one␈α∞section␈α∞of
␈↓ α←␈↓code␈α∞to␈α∞another␈α∞that␈α∞is␈α∂characterized␈α∞by␈α∞the␈α∞calling␈α∞of␈α∞procedures.␈α∂ This␈α∞is
␈↓ α←␈↓typically␈α
done␈α
at␈α
a␈α
selected␈α∞time␈α
and␈α
in␈α
a␈α
particular␈α∞context--both␈α
carefully
␈↓ α←␈↓chosen␈α↔by␈α↔the␈α↔programmer.␈α↔ If␈α↔a␈α↔no-op␈α↔is␈α↔substituted␈α↔for␈α↔a␈α↔missing
␈↓ α←␈↓␈↓2-4␈↓ λ∨PRODUCTION RULES    25␈↓

␈↓"β␈↓ α←␈↓procedure,␈α∂the␈α∞environment␈α∂upon␈α∂return␈α∞will␈α∂not␈α∞be␈α∂what␈α∂the␈α∞programmer
␈↓ α←␈↓expected,␈α⊃and␈α⊃subsequent␈α⊃procedure␈α⊃calls␈α⊃will␈α⊃be␈α⊃executed␈α⊃in␈α⊃increasingly
␈↓ α←␈↓incorrect␈αenvironments.␈α
Similarly,␈αprocedures␈α
that␈αhave␈α
been␈αadded␈α
must␈αbe
␈↓ α←␈↓called␈α∞from␈α∞␈↓↓somewhere␈↓␈α
in␈α∞the␈α∞program,␈α
but␈α∞the␈α∞location␈α
of␈α∞the␈α∞call␈α∞must␈α
be
␈↓ α←␈↓chosen carefully if the effect is to be meaningful.
␈↓"β␈↓ α←␈↓␈↓ β?Production␈α
systems,␈α
on␈α
the␈α∞other␈α
hand,␈α
emphasize␈α
the␈α∞decoupling␈α
of
␈↓ α←␈↓control␈α
flow␈α
from␈αthe␈α
writing␈α
of␈αrules.␈α
 Each␈α
rule␈α
is␈αdesigned␈α
to␈α
be␈αa␈α
modular
␈↓ α←␈↓chunk␈α∞of␈α∞knowledge␈α∞with␈α∞its␈α∞own␈α∞statement␈α∞of␈α∞relevance.␈α∞ Thus,␈α∞where␈α∞the
␈↓ α←␈↓␈↓¬ALGOL␈↓␈α∞programmer␈α∞carefully␈α∞chooses␈α∞the␈α∞order␈α∞of␈α∞procedure␈α∞calls␈α∞to␈α∞create␈α∞a
␈↓ α←␈↓selected␈α→sequence␈α→of␈α→environments,␈α~in␈α→a␈α→production␈α→system␈α→it␈α~is␈α→the
␈↓ α←␈↓environment␈α
which␈α
chooses␈α
the␈α
next␈α
rule␈α
for␈α
execution.␈α
 And␈α
since␈α
a␈αrule␈α
can
␈↓ α←␈↓only␈α⊂be␈α⊂chosen␈α⊃if␈α⊂its␈α⊂criteria␈α⊃of␈α⊂relevance␈α⊂have␈α⊃been␈α⊂met,␈α⊂the␈α⊃choice␈α⊂will
␈↓ α←␈↓continue␈αto␈α
be␈αa␈αplausible␈α
one,␈αand␈αsystem␈α
behavior␈αwill␈αremain␈α
``reasonable,''
␈↓ α←␈↓even as rules are successively deleted.

␈↓"β␈↓ α←␈↓␈↓α2-4-4    Production rules as a high-level language␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈α⊂noted,␈α⊂the␈α⊂inference␈α⊂rules␈α⊂are␈α⊂composed␈α⊂of␈α⊂clauses␈α⊂made␈α⊂up␈α⊂of
␈↓ α←␈↓predicate␈α∞functions␈α∞and␈α∞attribute-object-value␈α∞triples.␈α∞ The␈α∂entire␈α∞collection
␈↓ α←␈↓of␈αthese␈αelements␈αforms␈αa␈αset␈αof␈αconceptual␈αprimitives␈αfor␈αany␈αgiven␈αdomain.
␈↓ α←␈↓Thus␈α⊂in␈α⊂dealing␈α⊂with␈α⊂infectious␈α⊂disease,␈α⊂for␈α⊂example,␈α⊂there␈α⊂are␈α⊂␈↓	cultures␈↓
␈↓ α←␈↓with ␈↓	sites␈↓ like ␈↓	blood␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?If␈α∂we␈α∂consider␈α∂pushing␈α∂this␈α⊂back␈α∂one␈α∂level,␈α∂it␈α∂becomes␈α⊂plausible␈α∂to
␈↓ α←␈↓consider␈α↔the␈α↔idea␈α↔of␈α↔␈↓↓predicate␈α↔function,␈α↔attribute,␈α↔object,␈↓␈α↔and␈↓↓␈α↔value␈↓␈α↔as
␈↓ α←␈↓conceptual␈α!primitives␈α!in␈α!the␈α more␈α!general␈α!domain␈α!of␈α knowledge
␈↓ α←␈↓representation.␈α∂ We␈α∞might␈α∂consider␈α∞each␈α∂of␈α∞them␈α∂an␈α∞indication␈α∂of␈α∂a␈α∞whole
␈↓ α←␈↓class␈α⊗of␈α⊗objects,␈α⊗with␈α↔individual␈α⊗instances␈α⊗supplied␈α⊗by␈α⊗the␈α↔domain␈α⊗of
␈↓ α←␈↓application.␈α
 This␈α
suggests␈α
treating␈α
them␈α
as␈α
extended␈α
data␈α
types,␈α
which␈α
is␈αa
␈↓ α←␈↓useful␈α
analogy␈α
to␈α
keep␈αin␈α
mind.␈α
 There␈α
are␈α
13␈αsuch␈α
``data␈α
types,''␈α
used␈α
as␈αa␈α
set
␈↓ α←␈↓of␈α∂conceptual␈α∂primitives␈α∞for␈α∂expressing␈α∂knowledge␈α∞in␈α∂rule␈α∂form.␈α∂ They␈α∞are
␈↓ α←␈↓the␈αdata␈αtypes␈αof␈αwhat␈αis,␈αin␈αeffect,␈αa␈αhigh-level␈αprogramming␈αlanguage--one
␈↓ α←␈↓whose␈αsyntax␈α
is␈αvery␈αrestricted␈α
and␈αwhose␈α
sole␈αstatement␈αtype␈α
is␈αa␈αrule.␈α
 Since
␈↓ α←␈↓we␈α
refer␈α
to␈αthem␈α
often␈α
in␈αwhat␈α
follows,␈α
they␈αare␈α
listed␈α
and␈α
described␈αbelow,
␈↓ α←␈↓for reference:
␈↓ α←␈↓␈↓26    BACKGROUND␈↓ 
#2-4␈↓


␈↓"β␈↓ α←␈↓	predicate function
␈↓"β␈↓ α←␈↓	attribute              (abbreviated as ATTRIB)
␈↓"β␈↓ α←␈↓	object                 (for historical reasons, also called
␈↓"β␈↓ α←␈↓	                        a "context," abbreviated as CNTXT)
␈↓"β␈↓ α←␈↓	value
␈↓"β␈↓ α←␈↓	certainty factor       (a real number between -1 and 1, often
␈↓"β␈↓ α←␈↓	                        abbreviated CF)
␈↓"β␈↓ α←␈↓	list                   (the standard ␈↓¬LISP␈↓	 concept)
␈↓"β␈↓ α←␈↓	atom                   (the standard ␈↓¬LISP␈↓	 concept)
␈↓"β␈↓ α←␈↓	string                 (the standard ␈↓¬LISP␈↓	 concept)
␈↓"β␈↓ α←␈↓	slotname               (discussed in chapter 6)
␈↓"β␈↓ α←␈↓	slotexpert             (discussed in chapter 6)
␈↓"β␈↓ α←␈↓	blank                  (discussed in chapter 6)
␈↓"β␈↓ α←␈↓	advice                 (discussed in chapter 6)
␈↓"β␈↓ α←␈↓	table                  (discussed in chapter 6)

␈↓"β␈↓ α←␈↓␈↓ β?This␈αconcept␈αof␈αa␈αhigh-level␈αlanguage␈αwith␈αa␈αrestricted␈αsyntax␈αand␈αa
␈↓ α←␈↓small␈αnumber␈αof␈αdata␈αtypes␈αforms␈αan␈αimportant␈αbase␈αunderlying␈αmany␈αof␈αthe
␈↓ α←␈↓the␈α∞capabilities␈α∞of␈α∞␈↓¬TEIRESIAS␈↓.␈α∞ Perhaps␈α∂the␈α∞most␈α∞fundamental␈α∞of␈α∞these␈α∂is␈α∞the
␈↓ α←␈↓ability␈α∀to␈α∀make␈α∀multiple␈α∀uses␈α∀of␈α∀a␈α∀single␈α∀body␈α∀of␈α∃knowledge.␈α∀ Because
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓␈αcan␈αboth␈αassemble␈αand␈αdissect␈αrules,␈αwe␈αhave␈αa␈αsystem␈αthat␈α
can␈αnot
␈↓ α←␈↓only␈αuse␈αits␈αknowledge␈αof␈αa␈αdomain␈αdirectly,␈αbut␈αcan␈αalso␈αexamine␈αit,␈αalter␈αit,
␈↓ α←␈↓abstract␈α∞it,␈α∞and␈α∞draw␈α∞conclusions␈α∞about␈α∞it.␈α∞ The␈α∞rules␈α∞are␈α∞thus,␈α∞at␈α
different
␈↓ α←␈↓times,␈αboth␈αcode␈αand␈αdata␈αand␈αare␈αused␈αin␈αboth␈αcapacities␈αalmost␈αequally.␈α A
␈↓ α←␈↓large␈α⊂number␈α∂of␈α⊂interesting␈α∂and␈α⊂useful␈α∂features␈α⊂follow␈α∂from␈α⊂this;␈α⊂they␈α∂are
␈↓ α←␈↓explored␈α
in␈α
subsequent␈αchapters.␈α
 To␈α
clarify␈α
this␈αpoint,␈α
however,␈α
we␈α
offer␈αa
␈↓ α←␈↓simple but illustrative example.
␈↓"β␈↓ α←␈↓␈↓ β?As␈α
indicated␈αearlier␈α
in␈αour␈α
discussion␈αof␈α
the␈αcontrol␈α
structure,␈αbefore
␈↓ α←␈↓invoking␈α∂a␈α∂rule␈α∂the␈α∂system␈α∂performs␈α∞a␈α∂partial␈α∂evaluation␈α∂of␈α∂its␈α∂premise␈α∞to
␈↓ α←␈↓make␈α∞sure␈α∂that␈α∞the␈α∞rule␈α∂is␈α∞not␈α∞already␈α∂guaranteed␈α∞to␈α∞fail.␈α∂ But␈α∞performing
␈↓ α←␈↓this␈αevaluation␈αis␈αnontrivial.␈α The␈αsystem␈αrequires␈αa␈αway␈αto␈αtell␈αif␈αany␈αclause
␈↓ α←␈↓in␈αthe␈αpremise␈αis␈αknown␈αto␈αbe␈αfalse.␈α It␈αcannot␈αsimply␈αEVALuate␈αeach␈αclause
␈↓ α←␈↓individually,␈α
since␈α
a␈α
subgoal␈α
that␈α
had␈α
never␈α
been␈α
traced␈α
before␈α
would␈αsend
␈↓ α←␈↓the system off on its recursive search.
␈↓"β␈↓ α←␈↓␈↓ β?However,␈α∂if␈α∞the␈α∂system␈α∂can␈α∞establish␈α∂which␈α∂attribute␈α∞is␈α∂used␈α∂in␈α∞the
␈↓ α←␈↓clause,␈α∩it␈α⊃is␈α∩possible␈α⊃to␈α∩determine␈α⊃whether␈α∩this␈α⊃attribute␈α∩has␈α∩been␈α⊃traced
␈↓ α←␈↓previously␈α_(by␈α_reference␈α_to␈α_internal␈α→flags).␈α_ If␈α_so,␈α_the␈α_clause␈α→can␈α_be
␈↓ α←␈↓EVALuated to obtain the value.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∂process␈α⊂is␈α∂made␈α⊂possible␈α∂by␈α∂a␈α⊂␈↓	TEMPLATE␈↓␈α∂associated␈α⊂with␈α∂each
␈↓ α←␈↓predicate␈α∩function.␈α∩ It␈α∩describes␈α∩the␈α∩format␈α∩of␈α∩any␈α∩call␈α∩of␈α∩that␈α∩predicate
␈↓ α←␈↓function,␈α∩by␈α∩giving␈α∩the␈α∩generic␈α∩type␈α∩and␈α∩order␈α∩of␈α∩the␈α∩arguments␈α∪to␈α∩the
␈↓ α←␈↓function. It thus resembles a simplified procedure declaration.
␈↓ α←␈↓␈↓2-4␈↓ λ∨PRODUCTION RULES    27␈↓


␈↓"β␈↓ α←␈↓	Function           Template             Sample function call

␈↓"β␈↓ α←␈↓	 SAME      (SAME CNTXT ATTRIB VALUE)   (SAME CNTXT SITE BLOOD)


␈↓"β␈↓ α←␈↓α␈↓ ∧≠Fig. 2-7.    Example of a function template.    

␈↓ α←␈↓The␈αtemplate␈αis␈αnot␈αitself␈αa␈αpiece␈αof␈αcode␈αbut␈αis␈αsimply␈αa␈αlist␈αstructure␈αof␈αthe
␈↓ α←␈↓sort␈α⊂shown␈α⊂above,␈α⊂indicating␈α⊂the␈α⊂appearance␈α⊂of␈α⊂an␈α⊂interpreted␈α⊂call␈α⊂to␈α∂the
␈↓ α←␈↓predicate␈αfunction.␈α
 Since␈αrules␈α
are␈αkept␈αin␈α
interpreted␈αform,␈α
as␈αshown␈αin␈α
Fig.
␈↓ α←␈↓2-4,␈αthe␈αtemplate␈αcan␈αbe␈αused␈αas␈αa␈α
guide␈αto␈αdissect␈αa␈αrule.␈α For␈αeach␈αclause␈α
in
␈↓ α←␈↓a␈αrule,␈α
␈↓¬TEIRESIAS␈↓␈αretrieves␈α
the␈αtemplate␈αassociated␈α
with␈αthe␈α
predicate␈αfunction
␈↓ α←␈↓found␈αin␈αthat␈αclause␈α(i.e.,␈αthe␈αtemplate␈αassociated␈αwith␈αthe␈α␈↓	CAR␈↓␈αof␈αthe␈αclause),
␈↓ α←␈↓and␈α∩uses␈α∩it␈α∩to␈α∩guide␈α∩the␈α∩examination␈α∩of␈α∩the␈α∩clause.␈α∩ In␈α∩the␈α∩case␈α∩of␈α⊃the
␈↓ α←␈↓function␈α␈↓	SAME␈↓,␈αfor␈αinstance,␈αthe␈αtemplate␈αindicates␈αthat␈αthe␈αattribute␈α
(␈↓	ATTRIB␈↓)
␈↓ α←␈↓is␈α
the␈αthird␈α
element␈αof␈α
the␈αlist␈α
structure␈αthat␈α
comprises␈αthe␈α
function␈αcall.␈α
 The
␈↓ α←␈↓previewing␈α∂mechanism␈α∂uses␈α∂the␈α∂templates␈α∞to␈α∂extract␈α∂the␈α∂attribute␈α∂from␈α∞the
␈↓ α←␈↓clause in question and can then determine whether or not it has been traced.
␈↓"β␈↓ α←␈↓␈↓ β?There are two points of interest here:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?Part␈α⊃of␈α⊃the␈α⊃system␈α⊃is␈α⊃examining␈α⊃the␈α⊃code␈α⊃(the␈α⊃rules)␈α⊂being
␈↓ α←␈↓␈↓ β?executed by another part, and
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?this␈α∀examination␈α∀is␈α∀guided␈α∀by␈α∀the␈α∀information␈α∃carried␈α∀in
␈↓ α←␈↓␈↓ β?components of the rules themselves.

␈↓ α←␈↓The␈αability␈αto␈αexamine␈αthe␈αcode␈αcould␈αhave␈αbeen␈αaccomplished␈αby␈αrequiring
␈↓ α←␈↓all␈αpredicate␈α
functions␈αto␈α
use␈αthe␈α
same␈αformat,␈α
but␈αthis␈α
is␈αobviously␈α
awkward.
␈↓ α←␈↓Allowing␈α
each␈α
function␈αto␈α
describe␈α
the␈αformat␈α
of␈α
its␈αown␈α
calls␈α
permits␈αcode␈α
to
␈↓ α←␈↓be␈α
stylized␈αwithout␈α
being␈αconstrained␈α
to␈αa␈α
single␈αform,␈α
and␈αhence␈α
is␈αflexible
␈↓ α←␈↓and␈α⊂much␈α⊂easier␈α⊂to␈α⊂use.␈α⊂ This␈α⊂approach␈α⊂requires␈α⊂only␈α⊂that␈α⊂each␈α⊃form␈α⊂be
␈↓ α←␈↓expressible␈αin␈αa␈αtemplate␈αbuilt␈α
from␈αthe␈αcurrent␈αset␈αof␈α
conceptual␈αprimitives.
␈↓ α←␈↓It␈α
also␈α
insures␈α
that␈α
the␈α
capability␈α
will␈α
persist␈α
in␈α
the␈α
face␈α
of␈α
future␈α
additions␈α
to
␈↓ α←␈↓the␈α
system.␈α
The␈α
result␈α
is␈α
one␈α
example␈α
of␈α
the␈α
general␈α
idea␈α
of␈α
giving␈αthe␈α
system
␈↓ α←␈↓access␈α⊃to,␈α⊃and␈α⊃an␈α⊃``understanding''␈α⊃of,␈α⊃its␈α⊃own␈α⊃representations.␈α⊂ Additional
␈↓ α←␈↓examples of this concept are spread throughout the remainder of this work.

␈↓"β␈↓ α←␈↓␈↓α2-5    LEVELS OF KNOWLEDGE␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αconcept␈αof␈αmeta-level␈αknowledge␈αintroduced␈αin␈αchapter␈α1␈αcan␈αbe
␈↓ α←␈↓defined␈α∪more␈α∪generally␈α∪as␈α∀multiple␈α∪levels␈α∪of␈α∪knowledge.␈↓
5␈↓␈α∪This␈α∀idea␈α∪has
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[5]␈α∩There␈α⊃have␈α∩been␈α⊃other␈α∩uses␈α⊃of␈α∩the␈α⊃term␈α∩``levels␈α⊃of␈α∩knowledge,''␈α⊃most
␈↓ α←␈↓notably␈α⊃to␈α⊃describe␈α∩the␈α⊃hierarchy␈α⊃of␈α⊃domain␈α∩knowledge␈α⊃in␈α⊃the␈α∩␈↓¬HEARSAY␈α⊃II␈↓
␈↓ α←␈↓system␈α∞[Lesser74]␈α∞(e.g.,␈α∞phoneme,␈α∞word,␈α∞syntax,␈α∞semantics,␈α∞etc.).␈α∞ We␈α∂use␈α∞the
␈↓ α←␈↓term␈αhere␈α
in␈αa␈α
different␈αsense,␈αthat␈α
of␈αa␈α
hierarchy␈αof␈α
␈↓↓types␈↓␈αof␈αknowledge,␈α
and
␈↓ α←␈↓intend that meaning throughout.
␈↓ α←␈↓␈↓28    BACKGROUND␈↓ 
#2-5␈↓

␈↓"β␈↓ α←␈↓several␈α∩important␈α∩applications␈α∪in␈α∩the␈α∩work␈α∪reported␈α∩here.␈α∩ In␈α∪chapter␈α∩7,
␈↓ α←␈↓strategies␈α∞are␈α∞defined␈α∞in␈α∞terms␈α∂of␈α∞a␈α∞knowledge␈α∞hierarchy␈α∞with␈α∂an␈α∞arbitrary
␈↓ α←␈↓number␈α∞of␈α∞levels.␈α∞ A␈α∞different␈α∞sort␈α
of␈α∞hierarchy␈α∞is␈α∞responsible␈α∞for␈α∞much␈α
of
␈↓ α←␈↓the␈αperformance␈αand␈α
generality␈αof␈α␈↓¬TEIRESIAS␈↓'s␈αknowledge␈α
acquisition␈αroutines.
␈↓ α←␈↓In␈α∞this␈α∞hierarchy,␈α∞knowledge␈α∞is␈α∞stratified␈α∞into␈α∞three␈α∞distinct␈α∞levels.␈α∞ A␈α∞brief
␈↓ α←␈↓description␈αof␈αit␈αwill␈αhelp␈αto␈αclarify␈αmany␈αof␈αthe␈αideas␈αinvolved␈αand␈α
illustrate
␈↓ α←␈↓their power.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αfirst␈αlevel␈αof␈αthe␈αhierarchy␈αcontains␈αthe␈αobject-level␈αknowledge--
␈↓ α←␈↓medical␈α
knowledge␈α∞of␈α
cultures,␈α∞organisms,␈α
drugs,␈α∞etc.␈α
 High␈α∞performance␈α
on
␈↓ α←␈↓the␈α⊃task␈α⊃of␈α⊃diagnosis␈α⊃and␈α⊃therapy␈α⊃selection␈α⊃is␈α⊃supported␈α⊃by␈α∩an␈α⊃extensive
␈↓ α←␈↓collection␈αof␈αknowledge␈αabout␈αobjects␈αin␈αthe␈αmedical␈αdomain.␈αThis␈αfirst␈αlevel
␈↓ α←␈↓is naturally limited to this single domain.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂next␈α∂level␈α∂is␈α∂concerned␈α∂with␈α∂the␈α∂conceptual␈α∂building␈α∂blocks␈α∞of
␈↓ α←␈↓the␈α∃knowledge␈α⊗representation--the␈α∃predicate␈α∃functions,␈α⊗attributes,␈α∃values,
␈↓ α←␈↓rules,␈α∪and␈α∪so␈α∪on.␈α∪ Performance␈α∪on␈α∪the␈α∪task␈α∪of␈α∪knowledge␈α∀acquisition␈α∪is
␈↓ α←␈↓dependent␈α
upon␈α
an␈αextensive␈α
body␈α
of␈αknowledge␈α
about␈α
these␈αbuilding␈α
blocks.
␈↓ α←␈↓That␈α⊃is,␈α⊂there␈α⊃is␈α⊂assembled␈α⊃here␈α⊂a␈α⊃large␈α⊂amount␈α⊃of␈α⊂knowledge␈α⊃about␈α⊂the
␈↓ α←␈↓representational␈α∂primitives.␈α∂ As␈α∂will␈α∂become␈α∂clear␈α∂in␈α∂chapters␈α∂5␈α∂and␈α∂6,␈α∞the
␈↓ α←␈↓system␈α
has␈α
an␈α``understanding''␈α
of␈α
what␈αan␈α
attribute␈α
is,␈αwhat␈α
roles␈α
it␈αplays,␈α
etc.
␈↓ α←␈↓Since␈αno␈αreference␈αis␈αmade␈αat␈αthis␈αlevel␈αto␈αany␈αspecific␈αinstance␈αof␈αany␈αof␈αthe
␈↓ α←␈↓primitives,␈α⊂this␈α⊂level␈α⊂of␈α⊂knowledge␈α⊂has␈α⊂a␈α⊂degree␈α⊂of␈α⊂domain␈α∂independence.
␈↓ α←␈↓Over␈αthe␈αrange␈αof␈αdomains␈αin␈αwhich␈αknowledge␈αcan␈αbe␈αrepresented␈αin␈αterms
␈↓ α←␈↓of␈α
these␈α
primitives,␈α
the␈α
knowledge␈α
acquisition␈α
routines␈α
are␈α
similarly␈α
domain
␈↓ α←␈↓independent.
␈↓"β␈↓ α←␈↓␈↓ β?Knowledge␈α⊗at␈α⊗the␈α⊗third␈α↔level␈α⊗is␈α⊗concerned␈α⊗with␈α↔the␈α⊗conceptual
␈↓ α←␈↓primitives␈αbehind␈αrepresentations␈αin␈αgeneral.␈α To␈αmake␈αthis␈α
clearer,␈αconsider
␈↓ α←␈↓the␈α∪recursion␈α∪of␈α∪ideas: ␈α∪To␈α∩aid␈α∪the␈α∪construction␈α∪of␈α∪a␈α∪high␈α∩performance
␈↓ α←␈↓(object-level)␈α∀program,␈α∃we␈α∀build␈α∀a␈α∃(meta-level)␈α∀system␈α∀that␈α∃can␈α∀acquire
␈↓ α←␈↓object-level␈αknowledge.␈α
 Its␈αperformance␈αat␈α
this␈αtask␈αis␈α
based␈αon␈αan␈α
extensive
␈↓ α←␈↓store␈α∂of␈α∂knowledge␈α∂about␈α∂specific␈α⊂representations.␈α∂ ␈↓↓But␈α∂it␈α∂in␈α∂turn␈α⊂is␈α∂``just''
␈↓ α←␈↓↓another␈α"knowledge-based␈α"system.␈↓␈α"By␈α!supplying␈α"the␈α"proper␈α"set␈α!of
␈↓ α←␈↓representation␈α
independent␈α
primitives␈α
(and␈α
a␈α
store␈α
of␈α
knowledge␈αabout␈α
them),
␈↓ α←␈↓we␈αcan␈αuse␈α
precisely␈αthe␈αsame␈α
formalism␈α(indeed,␈αthe␈α
same␈αcode)␈αto␈αprovide␈α
a
␈↓ α←␈↓system for acquiring knowledge about individual representations.␈↓
6␈↓
␈↓"β␈↓ α←␈↓␈↓ β?In␈α_this␈α_way,␈α_the␈α_second-order␈α_system␈α_can␈α_be␈α_used␈α_to␈α_acquire
␈↓ α←␈↓knowledge␈αabout␈αa␈αrepresentation.␈α This␈αin␈αturn␈αbecomes␈αthe␈αknowledge␈α
base
␈↓ α←␈↓for␈α∪the␈α∪meta-level␈α∪system,␈α∪which␈α∪then␈α∪facilitates␈α∪the␈α∪construction␈α∀of␈α∪the
␈↓ α←␈↓knowledge␈α
base␈α
for␈αthe␈α
object-level␈α
performance␈αprogram.␈α
 The␈α
two␈αstages␈α
of
␈↓ α←␈↓this process are shown below.




␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[6]␈α⊃The␈α∩details␈α⊃and␈α⊃examples␈α∩of␈α⊃this␈α⊃bootstrapping␈α∩process␈α⊃are␈α∩given␈α⊃in
␈↓ α←␈↓chapter 6.
␈↓ α←␈↓%D2-5␈↓ εYLEVELS OF KNOWLEDGE    29%*



␈↓"␈↓ α←␈↓∧            teaching about        teaching about the
␈↓"␈↓ α←␈↓∧           a representation      domain of application
␈↓"␈↓ α←␈↓∧⊂ααααααααααααααα⊃      ⊂ααααααααααααααα⊃       ⊂αααααααααααααα⊃
␈↓"␈↓ α←␈↓∧~knowledge of   ~      ~knowledge of   ~       ~              ~
␈↓"␈↓ α←␈↓∧~representation-~      ~primitives for ~       ~ object-level ~
␈↓"␈↓ α←␈↓∧~independent    ~= = @ ~a specific     ~ = = @ ~knowledge base~
␈↓"␈↓ α←␈↓∧~primitives     ~      ~representation ~       ~              ~
␈↓"␈↓ α←␈↓∧%ααααααααααααααα$      %ααααααααααααααα$       %αααααααααααααα$

␈↓"␈↓ α←␈↓∧   SYSTEM 2                SYSTEM 1                SYSTEM 0


␈↓"␈↓ α←␈↓α␈↓ ∧ZFig. 2-8.    The conceptual process.    

␈↓ α←␈↓Note,␈α≠however,␈α≠that␈α≠while␈α≤this␈α≠indicates␈α≠the␈α≠process␈α≠as␈α≤it␈α≠appears
␈↓ α←␈↓conceptually,␈α∂a␈α∂more␈α∂accurate␈α∂system␈α⊂view␈α∂is␈α∂shown␈α∂below.␈α∂ Here␈α⊂we␈α∂have
␈↓ α←␈↓only␈α∞two␈α∞systems,␈α∞because␈α∞there␈α∞is␈α
in␈α∞fact␈α∞only␈α∞a␈α∞single␈α∞higher␈α∞level␈α
system.
␈↓ α←␈↓This␈α
is␈α
possible␈α
because␈α
the␈α
process␈α
of␈α
teaching␈α
about␈α
a␈α
␈↓↓representation␈↓␈αcan␈α
be
␈↓ α←␈↓made␈α∪computationally␈α∪identical␈α∪to␈α∪the␈α∪process␈α∪of␈α∪teaching␈α∪about␈α∩specific
␈↓ α←␈↓␈↓↓instances␈↓␈α∂of␈α∂that␈α∂representation.␈α∂ The␈α∂two␈α∂are␈α∂therefore␈α∂accomplished␈α∂with
␈↓ α←␈↓precisely the same mechanism.

␈↓"␈↓ α←␈↓∧                         teaching about the
␈↓"␈↓ α←␈↓∧                        domain of application
␈↓"␈↓ α←␈↓∧          ⊂ααααααααααααααα⊃               ⊂αααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧  ⊂αααα`≥ ~  meta-level   ~               ~  object-level  ~
␈↓"␈↓ α←␈↓∧  ~ ⊂αα≤' ~    system     ~    = = @      ~     system     ~
␈↓"␈↓ α←␈↓∧  ~ ~     %ααααααααααααααα$               %αααααααααααααααα$
␈↓"␈↓ α←␈↓∧  ~ ~            ~ ~
␈↓"␈↓ α←␈↓∧  ~ ~            ~ ~
␈↓"␈↓ α←␈↓∧  ~ %αααααααααααα$ ~
␈↓"␈↓ α←␈↓∧  %αααααααααααααααα$
␈↓"␈↓ α←␈↓∧    teaching about
␈↓"␈↓ α←␈↓∧   a representation

␈↓"␈↓ α←␈↓α␈↓ ∧AFig. 2-9.    The computational process.    

␈↓"β␈↓ α←␈↓␈↓ β?In␈α⊂␈↓¬TEIRESIAS␈↓,␈α⊂the␈α⊃recursive␈α⊂application␈α⊂of␈α⊂the␈α⊃knowledge␈α⊂acquisition
␈↓ α←␈↓system␈αoffers␈αa␈α
certain␈αdegree␈αof␈α
representation␈αindependence;␈αits␈α
extent␈αand
␈↓ α←␈↓limitations are examined in chapter 6.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α_began␈α_this␈α→report␈α_by␈α_noting␈α→that␈α_the␈α_original␈α→search␈α_for
␈↓ α←␈↓generality--set␈α⊗in␈α∃the␈α⊗domain␈α∃of␈α⊗problem-solving␈α⊗methods--has␈α∃proved
␈↓ α←␈↓unsuccessful␈αthus␈αfar.␈α Knowledge-based␈αmethods␈αhave␈αbeen␈αsuggested␈αas␈αan
␈↓ α←␈↓alternative,␈αbut␈αeach␈αhas␈αa␈αsharply␈αlimited␈αrange␈αof␈αapplication,␈αand␈αthe␈αlure
␈↓ α←␈↓of generality remains.  Is there another way to salvage it?
␈↓"β␈↓ α←␈↓␈↓ β?One␈α
of␈αthe␈α
underlying␈αthemes␈α
of␈αthis␈α
work␈αis␈α
the␈αattempt␈α
to␈αcapture␈α
a
␈↓ α←␈↓different␈αform␈αof␈αgenerality,␈αone␈αthat␈αhas␈αits␈αsource␈αin␈αknowledge␈αacquisition
␈↓ α←␈↓rather␈αthan␈αin␈αproblem-solving␈αmethods.␈α That␈αis,␈αif␈αprograms␈α
require␈αlarge
␈↓ α←␈↓␈↓30    BACKGROUND␈↓ 
#2-5␈↓

␈↓"β␈↓ α←␈↓stores␈α
of␈α
knowledge␈α
for␈α
performance,␈α
then␈α
can␈α
we␈α
not␈α
take␈α
a␈α
step␈α
back␈αand
␈↓ α←␈↓discover␈α⊗powerful,␈α↔broadly␈α⊗applicable␈α↔techniques␈α⊗for␈α↔accomplishing␈α⊗the
␈↓ α←␈↓transfer␈αof␈αknowledge␈αfrom␈αexpert␈αto␈αprogram?␈α The␈αresulting␈αman-machine
␈↓ α←␈↓combination␈αwould␈α
be␈αa␈αsemi-automatic␈α
system,␈αwhose␈αgenerality␈α
arose␈αfrom
␈↓ α←␈↓access␈αto␈αthe␈αappropriate␈αhuman␈αexperts␈αand␈αwhose␈αpower␈αwas␈αbased␈αon␈αthe
␈↓ α←␈↓store␈α∂of␈α∞knowledge␈α∂it␈α∞acquired␈α∂from␈α∞them.␈α∂ The␈α∞next␈α∂five␈α∂chapters␈α∞discuss
␈↓ α←␈↓several steps toward the realization of this aim.
␈↓ α←␈↓␈↓␈↓ 
⊃    31␈↓




␈↓"β␈↓ α←␈↓␈↓ ε∃␈↓αChapter 3



␈↓"β␈↓ α←␈↓α␈↓ ∧f␈↓λEXPLANATION␈↓α


␈↓"β␈↓ α←␈↓α␈↓ ¬$how to give an accounting









␈↓"β␈↓ α←␈↓␈↓ ¬GWhat␈α∂do␈α⊂you␈α∂mean?␈α∂ What␈α⊂you␈α∂have␈α∂said␈α⊂so␈α∂far
␈↓ α←␈↓␈↓ ¬Gleaves me uncertain whether to trust or fear.
␈↓"β␈↓ α←␈↓␈↓ ∧␈↓↓Oedipus the King␈↓, lines 88-90␈↓↓);BEGINSMALLQUOTE;SKIP 1;
␈↓"β␈↓ α←␈↓␈↓ β?``No,␈α⊃no!␈α⊂The␈α⊃adventures␈α⊂first,''␈α⊃said␈α⊂the␈α⊃Gryphon␈α⊂in␈α⊃an␈α⊂impatient
␈↓ α←␈↓tone: ``explanations take such a dreadfully long time!''
␈↓"β␈↓ α←␈↓␈↓ 	A[Carroll60]

␈↓"β␈↓ α←␈↓␈↓α3-1    INTRODUCTION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∀fundamental␈α∀goal␈α∃of␈α∀an␈α∀explanation␈α∃facility␈α∀is␈α∀to␈α∃enable␈α∀a
␈↓ α←␈↓program␈αto␈αdisplay␈αa␈αcomprehensible␈αaccount␈αof␈αthe␈αmotivation␈αfor␈αall␈αof␈αits
␈↓ α←␈↓actions.␈α∞ This␈α∞chapter␈α∞explores␈α∞steps␈α∞taken␈α∞toward␈α∞this␈α∞goal,␈α∞examining␈α
the
␈↓ α←␈↓extent␈α∂to␈α∂which␈α⊂``all''␈α∂actions␈α∂of␈α∂a␈α⊂program␈α∂can␈α∂be␈α∂explained.␈α⊂ It␈α∂considers
␈↓ α←␈↓what␈α
is␈α
required␈α
for␈α
a␈α
``comprehensible''␈α
account␈α
and␈α
offers␈α
a␈α
framework␈αin
␈↓ α←␈↓which to understand ``motivations'' for one particular system.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∞utility␈α∞of␈α∂an␈α∞explanation␈α∞facility␈α∞should␈α∂be␈α∞clear.␈α∞ Even␈α∂for␈α∞an
␈↓ α←␈↓experienced␈α∂programmer,␈α⊂the␈α∂attempt␈α∂to␈α⊂account␈α∂for␈α∂program␈α⊂behavior␈α∂by
␈↓ α←␈↓hand␈α⊂simulation␈α⊃is␈α⊂difficult␈α⊂for␈α⊃any␈α⊂sizable␈α⊂program.␈α⊃ It␈α⊂is␈α⊃often␈α⊂difficult
␈↓ α←␈↓enough␈αto␈αdiscover␈αhow␈αa␈αprogram␈αgot␈αto␈αwhere␈αit␈αis.␈α Trying␈αto␈αaccount␈αfor
␈↓ α←␈↓past␈α∀behavior␈α∀(e.g.,␈α∀function␈α∀calls␈α∀that␈α∀have␈α∀long␈α∀since␈α∀exited)␈α∃is␈α∀often
␈↓ α←␈↓impossible because critical variables have been overwritten.
␈↓"β␈↓ α←␈↓␈↓ β?For␈α
consultation␈α
programs,␈α
in␈α
particular,␈α
the␈α
problem␈α
of␈α
explanation
␈↓ α←␈↓is␈αworse␈αbecause␈αthey␈αdeal␈αwith␈αan␈αaudience␈αassumed␈αto␈αknow␈αnothing␈αabout
␈↓ α←␈↓programming.␈α∞ This␈α
requires␈α∞a␈α
different␈α∞standard␈α
of␈α∞comprehensibility,␈α
one
␈↓ α←␈↓defined␈α⊃in␈α∩terms␈α⊃of␈α⊃the␈α∩application␈α⊃domain␈α⊃rather␈α∩than␈α⊃the␈α∩language␈α⊃of
␈↓ α←␈↓computation.␈α⊃ A␈α⊃naive␈α⊃user␈α⊃(e.g.,␈α⊃a␈α⊃student)␈α⊃should␈α⊃find␈α⊃the␈α⊃explanations
␈↓ α←␈↓␈↓32    EXPLANATION␈↓ 
#3-1␈↓

␈↓"β␈↓ α←␈↓educational;␈αa␈αmore␈α
experienced␈αuser␈αshould␈α
find␈αthem␈αreassuring,␈α
employing
␈↓ α←␈↓them␈αto␈αsatisfy␈αhimself␈αthat␈αthe␈αconclusions␈αthe␈αsystem␈αhas␈αreached␈αare␈αbased
␈↓ α←␈↓on␈α↔acceptable␈α⊗reasoning;␈α↔while␈α↔an␈α⊗expert␈α↔should␈α⊗find␈α↔them␈α↔useful␈α⊗in
␈↓ α←␈↓discovering␈α∞gaps␈α∞or␈α∞errors␈α∞in␈α
the␈α∞knowledge␈α∞base.␈α∞ We␈α∞describe␈α∞below␈α
how
␈↓ α←␈↓each of these objectives is accomplished.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∂chapter␈α⊂begins␈α∂by␈α⊂outlining␈α∂and␈α⊂discussing␈α∂the␈α⊂plausibility␈α∂of
␈↓ α←␈↓the␈α
fundamental␈αassumptions␈α
behind␈αthe␈α
techniques␈αused␈α
in␈α
␈↓¬TEIRESIAS␈↓.␈α This
␈↓ α←␈↓is␈αfollowed␈αby␈αseveral␈αexamples␈αof␈αthe␈αcapabilities␈αthat␈αhave␈αbeen␈αdeveloped
␈↓ α←␈↓and␈α∀a␈α∀description␈α∀of␈α∀how␈α∃they␈α∀are␈α∀achieved.␈α∀ Finally,␈α∀as␈α∀a␈α∃prelude␈α∀to
␈↓ α←␈↓knowledge␈α∂acquisition,␈α∂we␈α∂end␈α∂by␈α∂exploring␈α∂how␈α∂␈↓¬TEIRESIAS␈↓␈α∂may␈α∂be␈α∂used␈α∂to
␈↓ α←␈↓discover the source of problems in the knowledge base.

␈↓"β␈↓ α←␈↓␈↓α3-2    BASIC ASSUMPTIONS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∪techniques␈α∪used␈α∀in␈α∪␈↓¬TEIRESIAS␈↓␈α∪for␈α∪generating␈α∀explanations␈α∪are
␈↓ α←␈↓based␈αon␈αa␈αnumber␈αof␈αassumptions␈αabout␈αthe␈αsystem␈αbeing␈αexplained.␈α
 These
␈↓ α←␈↓assumptions␈α↔are␈α⊗reviewed␈α↔here␈α⊗to␈α↔help␈α⊗motivate␈α↔what␈α⊗follows␈α↔and␈α⊗to
␈↓ α←␈↓characterize their range of applicability.

␈↓"β␈↓ α←␈↓␈↓α3-2-1    Generalities:  Two assumptions␈↓
␈↓"β␈↓ α←␈↓␈↓ β?We␈αassume,␈αfirst,␈αthat␈αa␈αrecap␈αof␈αprogram␈αactions␈αcan␈αbe␈αan␈αeffective
␈↓ α←␈↓explanation␈αas␈αlong␈αas␈αthe␈αcorrect␈αlevel␈αof␈αdetail␈αis␈αchosen.␈α This␈αassumption
␈↓ α←␈↓simplifies␈αthe␈αtask␈αconsiderably,␈αsince␈α
it␈αmeans␈αthat␈αthe␈αsolution␈αrequires␈α
only
␈↓ α←␈↓the␈αability␈αto␈αrecord␈αand␈αplay␈αback␈αa␈αhistory␈αof␈αevents.␈α In␈αparticular,␈αit␈αrules
␈↓ α←␈↓out any need to simplify those events.
␈↓"β␈↓ α←␈↓␈↓ β?But␈α⊂this␈α∂assumption␈α⊂is␈α⊂perhaps␈α∂the␈α⊂source␈α∂of␈α⊂greatest␈α⊂limitation␈α∂as
␈↓ α←␈↓well.␈α
 It␈α
is␈α∞not␈α
obvious,␈α
for␈α∞instance,␈α
that␈α
an␈α∞appropriate␈α
level␈α
of␈α∞detail␈α
can
␈↓ α←␈↓always␈αbe␈αfound.␈α A␈αlarge␈αprogram␈αwith␈αcooperating␈αparallel␈αprocesses␈αmight
␈↓ α←␈↓prove␈αsufficiently␈αcomplex␈αthat␈αit␈αrequired␈αa␈αsophisticated␈αinterpretation␈αand
␈↓ α←␈↓simplification␈αto␈α
be␈αcomprehensible.␈α
 Neither␈αis␈α
it␈αobvious␈α
how␈αthis␈α
approach
␈↓ α←␈↓can␈α∂be␈α⊂applied␈α∂to␈α⊂programs␈α∂that␈α⊂are␈α∂primarily␈α⊂numeric.␈α∂ With␈α⊂a␈α∂program
␈↓ α←␈↓that␈α≠does␈α≠symbolic␈α≠reasoning,␈α≠recapping␈α≠offers␈α≠an␈α≠easily␈α≠understood
␈↓ α←␈↓explanation.␈α∂ But␈α∂simply␈α∂recapping␈α∞the␈α∂arithmetic␈α∂involved␈α∂in␈α∞determining
␈↓ α←␈↓parameters␈αof␈αa␈αcomplex␈αelectrical␈αnetwork,␈αfor␈αexample,␈αwould␈αexplain␈αlittle
␈↓ α←␈↓of␈α∞the␈α
reasoning␈α∞involved␈α∞and␈α
would␈α∞teach␈α
little␈α∞physics.␈α∞ Understanding␈α
it
␈↓ α←␈↓requires␈α∞a␈α∂much␈α∞higher␈α∞level␈α∂of␈α∞sophistication: ␈α∞It␈α∂assumes␈α∞that␈α∂the␈α∞viewer
␈↓ α←␈↓can␈α⊗interpret␈α⊗each␈α⊗numeric␈α⊗step␈α⊗in␈α⊗symbolic␈α⊗terms.␈α⊗ The␈α⊗lack␈α↔of␈α⊗any
␈↓ α←␈↓mechanism␈α∞for␈α∞either␈α∞simplifying␈α∞or␈α∞reinterpreting␈α∞computations␈α∞means␈α
our
␈↓ α←␈↓approach␈α∃is␈α∃basically␈α∃a␈α∃first␈α∃order␈α∃solution␈α∃to␈α∃the␈α∃general␈α⊗problem␈α∃of
␈↓ α←␈↓explaining program behavior.
␈↓"β␈↓ α←␈↓␈↓ β?If␈α∩a␈α∪simple␈α∩recap␈α∩is␈α∪going␈α∩to␈α∩be␈α∪effective,␈α∩there␈α∩must␈α∪be␈α∩several
␈↓ α←␈↓constraints␈αon␈αthe␈αlevel␈αof␈αdetail␈αchosen.␈α It␈αmust␈αbe␈α␈↓↓detailed␈αenough␈↓␈αthat␈αthe
␈↓ α←␈↓operations␈αit␈αcites␈αare␈αcomprehensible.␈α For␈αexample,␈αif␈αa␈αchess␈αprogram␈αwere
␈↓ α←␈↓to␈α
explain␈αa␈α
move␈αwith␈α
the␈αjustification␈α
that␈αit␈α
``picked␈αthe␈α
best␈α
choice,''␈αthe
␈↓ α←␈↓explanation␈α⊂would␈α∂explain␈α⊂very␈α⊂little␈α∂because␈α⊂it␈α∂wouldn't␈α⊂reveal␈α⊂what␈α∂was
␈↓ α←␈↓involved␈α∃in␈α∃the␈α∃operation␈α∃of␈α∃choosing.␈α∃ Some␈α∃explanation␈α∃in␈α⊗terms␈α∃of
␈↓ α←␈↓␈↓3-2␈↓ λ⊗BASIC ASSUMPTIONS    33␈↓

␈↓"β␈↓ α←␈↓alpha/beta␈α∃search␈α∃and␈α∃evaluation␈α∃functions␈α∃might␈α∃provide␈α⊗the␈α∃relevant
␈↓ α←␈↓information.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α→level␈α→must␈α→also␈α→be␈α_␈↓↓high␈α→enough␈↓␈α→that␈α→the␈α→operations␈α_are
␈↓ α←␈↓meaningful␈α⊗to␈α⊗the␈α⊗observer␈α↔and␈α⊗that␈α⊗unnecessary␈α⊗detail␈α↔is␈α⊗suppressed.
␈↓ α←␈↓Describing␈α⊃the␈α⊂chess␈α⊃program␈α⊂in␈α⊃terms␈α⊂of␈α⊃register-transfer␈α⊃operations,␈α⊂for
␈↓ α←␈↓instance,␈α∩would␈α∩lose␈α∩any␈α∩sense␈α∩of␈α∩task␈α∩specificity␈α∩and␈α∩introduce␈α∩pointless
␈↓ α←␈↓detail.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α
the␈α
explanation␈αmust␈α
be␈α
␈↓↓complete␈αenough␈↓␈α
that␈α
the␈αoperations
␈↓ α←␈↓cited␈α∂are␈α⊂sufficient␈α∂to␈α∂account␈α⊂for␈α∂all␈α∂behavior.␈α⊂ Completeness␈α∂is␈α⊂easiest␈α∂to
␈↓ α←␈↓achieve␈α∂if␈α⊂the␈α∂operations␈α∂are␈α⊂free␈α∂of␈α⊂side␈α∂effects␈α∂and␈α⊂the␈α∂system␈α⊂design␈α∂is
␈↓ α←␈↓reasonably␈α
``clean.'' ␈αIf␈α
the␈α
alpha/beta␈αsearch␈α
used␈α
in␈αthe␈α
chess␈α
program␈αhad
␈↓ α←␈↓numerous␈α
subtle␈α
side␈α
effects,␈α
it␈α
would␈αbe␈α
difficult␈α
to␈α
find␈α
a␈α
level␈α
of␈αdetail␈α
that
␈↓ α←␈↓could␈α⊗account␈α⊗for␈α∃the␈α⊗side␈α⊗effects␈α∃without␈α⊗introducing␈α⊗other␈α∃irrelevant
␈↓ α←␈↓information.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αsecond␈α
major␈αassumption␈α
is␈αthat␈α
there␈αexists␈α
some␈αframework␈α
for
␈↓ α←␈↓viewing␈α
the␈α∞program's␈α
actions␈α
that␈α∞will␈α
allow␈α
them␈α∞to␈α
be␈α∞comprehensible␈α
to
␈↓ α←␈↓the␈α∂observer.␈α∂ The␈α⊂likely␈α∂validity␈α∂of␈α∂this␈α⊂assumption␈α∂depends␈α∂on␈α⊂both␈α∂the
␈↓ α←␈↓program's␈α↔fundamental␈α_mechanisms␈α↔and␈α↔the␈α_level␈α↔at␈α↔which␈α_these␈α↔are
␈↓ α←␈↓examined.␈α
 Consider␈α
a␈α
program␈α
that␈α
does␈α
medical␈α
diagnosis␈α
using␈αa␈α
statistical
␈↓ α←␈↓approach␈α_based␈α↔on␈α_Bayes'␈α↔Theorem.␈α_ It␈α↔is␈α_difficult␈α↔to␈α_imagine␈α↔what
␈↓ α←␈↓explanation␈α∂of␈α∂its␈α∞actions␈α∂the␈α∂program␈α∞could␈α∂give␈α∂if␈α∞it␈α∂were␈α∂queried␈α∞while
␈↓ α←␈↓computing␈α∀probabilities.␈α∃ No␈α∀matter␈α∃what␈α∀level␈α∀of␈α∃detail␈α∀is␈α∃chosen,␈α∀the
␈↓ α←␈↓approach␈αis␈αnot␈α(nor␈αis␈αit␈αintended␈αto␈αbe)␈αan␈αaccurate␈αmodel␈αof␈αthe␈αreasoning
␈↓ α←␈↓process␈α≤typically␈α≤employed␈α≤by␈α≤physicians␈α≤(see␈α≤[Tversky74]␈α≤for␈α≠some
␈↓ α←␈↓experimental␈α
verification).␈α
 As␈α
effective␈α
as␈αthese␈α
actions␈α
may␈α
be,␈α
there␈α
is␈αno
␈↓ α←␈↓easy␈αway␈αto␈αinterpret␈αthem␈αin␈αterms␈αthat␈αwill␈αmake␈αthem␈αcomprehensible␈αto␈αa
␈↓ α←␈↓physician unacquainted with the program.
␈↓"β␈↓ α←␈↓␈↓ β?With␈αthe␈αcurrent␈αstate␈αof␈αthe␈α
art,␈αthen,␈αthe␈αdesire␈αto␈αhave␈α
a␈αprogram
␈↓ α←␈↓capable␈α
of␈αexplaining␈α
its␈αactions␈α
strongly␈αconstrains␈α
the␈α
methodology␈αchosen
␈↓ α←␈↓and␈α∞the␈α
control␈α∞structure␈α
that␈α∞can␈α
be␈α∞used.␈α
 There␈α∞do␈α
not␈α∞yet␈α
appear␈α∞to␈α
be
␈↓ α←␈↓general␈α
principles␈α
for␈α
generating␈αexplanations␈α
of␈α
arbitrary␈α
control␈αstructures
␈↓ α←␈↓in␈α_the␈α_way,␈α_for␈α_example,␈α_that␈α_an␈α_experienced␈α_programmer␈α_can␈α↔read
␈↓ α←␈↓unfamiliar␈α∀code␈α∪and␈α∀then␈α∪explain␈α∀it␈α∀to␈α∪someone␈α∀else.␈α∪ As␈α∀a␈α∀result,␈α∪the
␈↓ α←␈↓capability␈α⊃cannot␈α⊃now␈α⊃be␈α⊂tacked␈α⊃on␈α⊃to␈α⊃an␈α⊂existing␈α⊃system.␈α⊃ To␈α⊃make␈α⊂the
␈↓ α←␈↓problem␈α
tractable,␈αthe␈α
desired␈α
capabilities␈αmust␈α
be␈α
taken␈αinto␈α
account␈αearly␈α
in
␈↓ α←␈↓the system-design process.

␈↓"β␈↓ α←␈↓␈↓α3-2-2    Specifics:  How the assumptions were applied␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αfundamental␈αorganization␈αof␈αthe␈αperformance␈αprogram␈α
described
␈↓ α←␈↓earlier␈αprovides␈α
an␈αenvironment␈α
in␈αwhich␈α
both␈αof␈α
these␈αassumptions␈α
can␈αbe
␈↓ α←␈↓satisfied.␈α⊃ The␈α⊃simple␈α⊃and/or␈α⊂goal-tree␈α⊃control␈α⊃structure␈α⊃and␈α⊃the␈α⊂domain-
␈↓ α←␈↓specific␈αrules␈αinvoked␈αin␈αa␈α␈↓↓modus␈αponens␈↓␈αmode␈αoffer␈αa␈αbasis␈αfor␈α
explanations
␈↓ α←␈↓that␈αtypically␈αneed␈α
little␈αadditional␈αclarification.␈α The␈α
invocation␈αof␈αa␈α
rule␈αis
␈↓ α←␈↓taken␈α
as␈α
the␈αfundamental␈α
action␈α
of␈αthe␈α
system.␈α
 This,␈αalong␈α
with␈α
the␈αgoal␈α
tree
␈↓ α←␈↓as␈α⊂a␈α∂framework,␈α⊂accounts␈α∂for␈α⊂enough␈α⊂of␈α∂the␈α⊂system's␈α∂operation␈α⊂to␈α⊂make␈α∂a
␈↓ α←␈↓␈↓34    EXPLANATION␈↓ 
#3-2␈↓

␈↓"β␈↓ α←␈↓recap␈αof␈α
such␈αactions␈αan␈α
acceptable␈αexplanation.␈α In␈α
terms␈αof␈α
the␈αconstraints
␈↓ α←␈↓noted␈α∞earlier,␈α∞it␈α∂is␈α∞sufficiently␈α∞detailed--the␈α∞actions␈α∂performed␈α∞by␈α∞a␈α∂rule␈α∞in
␈↓ α←␈↓making␈αa␈α``conclusion,''␈αfor␈αinstance,␈αcorrespond␈αclosely␈αenough␈αto␈αthe␈αnormal
␈↓ α←␈↓connotation␈α
of␈α
this␈α
word␈α
that␈α
no␈α
greater␈α
detail␈α
is␈α
necessary.␈α
 It␈α
is␈α
still␈α
at␈α
an
␈↓ α←␈↓abstract␈α∀enough␈α∪level␈α∀that␈α∪the␈α∀operations␈α∪are␈α∀meaningful.␈α∪ Finally,␈α∀it␈α∪is
␈↓ α←␈↓generally␈α∃complete␈α∃enough--there␈α∃are␈α∃typically␈α∃no␈α∃other␈α∃mechanisms␈α∃or
␈↓ α←␈↓sources␈α↔of␈α↔information␈α↔that␈α↔the␈α↔observer␈α↔needs␈α↔to␈α↔know␈α↔in␈α_order␈α↔to
␈↓ α←␈↓understand how the program reaches its conclusions.␈↓
1␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
success␈αof␈α
this␈αtechnique␈α
relies␈α
to␈αsome␈α
extent␈αon␈α
the␈α
claim␈αthat
␈↓ α←␈↓the␈α∂performance␈α⊂program's␈α∂approach␈α⊂to␈α∂its␈α⊂domain␈α∂is␈α⊂sufficiently␈α∂intuitive
␈↓ α←␈↓that␈αa␈α
summary␈αof␈α
those␈αactions␈α
is␈αa␈α
reasonable␈αbasis␈α
for␈αexplanation.␈α
 While
␈↓ α←␈↓we␈αhave␈αnot␈αyet␈αattempted␈αto␈αprove␈αthe␈αclaim␈αin␈αany␈αformal␈αsense,␈α
there␈αare
␈↓ α←␈↓several factors that suggest its plausibility.
␈↓"β␈↓ α←␈↓␈↓ β?First,␈α∂the␈α∂performance␈α∂program␈α∂is␈α∂dealing␈α∂with␈α∂a␈α∂domain␈α∂in␈α∞which
␈↓ α←␈↓deduction,␈α
and␈α
deduction␈α∞in␈α
the␈α
face␈α
of␈α∞uncertainty,␈α
is␈α
a␈α
primary␈α∞task.␈α
The
␈↓ α←␈↓use␈α∂of␈α∂production␈α∂rules␈α∂seems␈α∂therefore␈α∂to␈α∂be␈α∂a␈α∂natural␈α∂way␈α∂of␈α∞expressing
␈↓ α←␈↓things␈α→about␈α→the␈α~domain␈α→and␈α→the␈α→display␈α~of␈α→such␈α→rules␈α~should␈α→be
␈↓ α←␈↓comprehensible.␈α Second,␈αthe␈αuse␈αof␈αsuch␈αrules␈αin␈αa␈αbackward-chaining␈α
mode
␈↓ α←␈↓seems␈α⊗to␈α⊗be␈α⊗a␈α⊗reasonably␈α∃intuitive␈α⊗scheme.␈α⊗ ␈↓↓Modus␈α⊗ponens␈↓␈α⊗is␈α⊗a␈α∃well-
␈↓ α←␈↓understood␈α
and␈αwidely␈α
(if␈αnot␈α
explicitly)␈αused␈α
mode␈αof␈α
inference.␈α
 Thus,␈αthe
␈↓ α←␈↓general␈αform␈αof␈αthe␈αrepresentation␈αand␈αthe␈αway␈αit␈αis␈αemployed␈αshould␈αnot␈αbe
␈↓ α←␈↓unfamiliar␈α∩to␈α∩the␈α∩average␈α∩user.␈α∩ More␈α∩specifically,␈α∩however,␈α∪consider␈α∩the
␈↓ α←␈↓source␈α⊗of␈α⊗the␈α⊗rules.␈α⊗ They␈α⊗are␈α⊗supplied␈α⊗by␈α⊗human␈α⊗experts␈α⊗who␈α∃were
␈↓ α←␈↓attempting␈α∂to␈α∞formalize␈α∂their␈α∞own␈α∂knowledge␈α∞of␈α∂the␈α∞domain.␈α∂ As␈α∂such,␈α∞the
␈↓ α←␈↓rules␈α∪embody␈α∪accepted␈α∪patterns␈α∪of␈α∪human␈α∪reasoning,␈α∪implying␈α∪that␈α∪they
␈↓ α←␈↓should␈αbe␈αrelatively␈αeasy␈αto␈αunderstand,␈αespecially␈αfor␈αthose␈αfamiliar␈αwith␈αthe
␈↓ α←␈↓domain.␈α As␈αsuch,␈αthey␈αalso␈αattack␈α
the␈αproblem␈αat␈αwhat␈αhas␈αbeen␈α
judged␈αan
␈↓ α←␈↓appropriate␈α
level␈α
of␈αdetail.␈α
 That␈α
is,␈αthey␈α
embody␈α
the␈αright␈α
size␈α
``chunks''␈αof
␈↓ α←␈↓the problem to be comprehensible.
␈↓"β␈↓ α←␈↓␈↓ β?Many␈α∃of␈α∃the␈α∀capabilities␈α∃of␈α∃the␈α∀current␈α∃explanation␈α∃system␈α∀also
␈↓ α←␈↓depend␈α∂on␈α⊂the␈α∂presence␈α⊂of␈α∂a␈α∂high-level␈α⊂language␈α∂of␈α⊂the␈α∂sort␈α⊂described␈α∂in
␈↓ α←␈↓chapter␈α2.␈α Extensive␈αuse␈αis␈αmade␈αof␈αthe␈αstylized␈αcode␈αand␈αthe␈αsmall␈αnumber
␈↓ α←␈↓of␈α⊃classes␈α∩of␈α⊃primitives␈α⊃found␈α∩in␈α⊃this␈α⊃language.␈α∩ This␈α⊃makes␈α∩possible,␈α⊃in
␈↓ α←␈↓particular,␈α
dissection␈α
and␈α
interpretation␈α∞of␈α
the␈α
rules,␈α
techniques␈α∞which␈α
form
␈↓ α←␈↓the basis for many of ␈↓¬TEIRESIAS␈↓'s capabilities.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α"have␈α!referred␈α"several␈α"times␈α!to␈α"explanations␈α"that␈α!are

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[1]␈α∞The␈α∞hedging␈α
here␈α∞arises␈α∞because,␈α
on␈α∞those␈α∞occasions␈α∞when␈α
explanations
␈↓ α←␈↓produced␈αby␈α␈↓¬TEIRESIAS␈↓␈αare␈αcryptic,␈αit␈αis␈αoften␈αa␈αresult␈αof␈αincompleteness␈αof␈α
this
␈↓ α←␈↓sort.␈α
 There␈α
are,␈α
for␈α
instance,␈α
rules␈α
in␈α
the␈α
performance␈α
program␈αwhose␈α
format
␈↓ α←␈↓and␈αcontent␈αhave␈αbeen␈αinfluenced␈αby␈αthe␈αattempt␈αto␈αtake␈αadvantage␈αof␈αsubtle
␈↓ α←␈↓aspects␈αof␈αthe␈αcontrol␈αstructure␈α(e.g.,␈αordering␈αthe␈αclauses␈αof␈αa␈αpremise␈αso␈αthat
␈↓ α←␈↓certain␈αattributes␈αare␈αtraced␈αfirst).␈α Since␈αthere␈αis␈αin␈αthe␈αsystem␈αno␈αindication
␈↓ α←␈↓of␈α⊗which␈α⊗rules␈α⊗have␈α⊗been␈α∃so␈α⊗modified,␈α⊗the␈α⊗simple␈α⊗goal-tree␈α⊗model␈α∃is
␈↓ α←␈↓incomplete is this respect.
␈↓ α←␈↓␈↓3-2␈↓ λ⊗BASIC ASSUMPTIONS    35␈↓

␈↓"β␈↓ α←␈↓``comprehensible''␈αand␈α``complete,''␈αwhich␈αraises␈αthe␈αquestions␈αComprehensible
␈↓ α←␈↓to␈αwhom?␈α and␈α
Complete␈αenough␈αfor␈αwhom? ␈α
As␈αindicated,␈αour␈α
efforts␈αhave
␈↓ α←␈↓been␈αdirected␈αat␈αusers␈αfrom␈αthe␈αapplication␈αdomain.␈α It␈αis␈αwith␈αrespect␈αto␈αthis
␈↓ α←␈↓audience␈α∪that␈α∪``comprehensible''␈α∀and␈α∪``complete''␈α∪are␈α∀used,␈α∪and␈α∪it␈α∀is␈α∪with
␈↓ α←␈↓respect␈α∃to␈α∃their␈α∀conceptual␈α∃level␈α∃that␈α∀appropriate␈α∃explanations␈α∃must␈α∀be
␈↓ α←␈↓produced.␈α
 While␈α∞a␈α
different␈α∞level␈α
would␈α
have␈α∞to␈α
be␈α∞chosen␈α
for␈α∞a␈α
different
␈↓ α←␈↓audience␈α
(e.g.,␈α
experienced␈α
programmers),␈α
the␈α
criteria␈α
above␈α
remain␈α∞valid␈α
if
␈↓ α←␈↓the explanations are to be based on a recap of program actions.

␈↓"β␈↓ α←␈↓␈↓α3-3    DESIGN CRITERIA␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈α⊃were␈α⊃three␈α⊃criteria␈α⊃central␈α⊂to␈α⊃the␈α⊃design␈α⊃of␈α⊃the␈α⊂explanation
␈↓ α←␈↓facilities.

␈↓"β␈↓ α←␈↓␈↓ ββ(1)␈↓ β?␈↓↓Accuracy␈↓.␈α∂ Above␈α∂all␈α∂else,␈α∂the␈α∂explanations␈α∂generated␈α⊂had␈α∂to
␈↓ α←␈↓␈↓ β?provide␈α∪an␈α∩accurate␈α∪picture␈α∩of␈α∪what␈α∩was␈α∪going␈α∩on␈α∪in␈α∩the
␈↓ α←␈↓␈↓ β?performance␈α≡program.␈α∨ This␈α≡meant␈α∨overcoming␈α≡several
␈↓ α←␈↓␈↓ β?temptations;␈αin␈αparticular,␈αthe␈αdesire␈αto␈α``dress␈αthings␈αup␈αjust␈αa
␈↓ α←␈↓␈↓ β?bit,''␈α↔to␈α⊗cover␈α↔over␈α↔some␈α⊗of␈α↔the␈α⊗less␈α↔impressive␈α↔(or␈α⊗less
␈↓ α←␈↓␈↓ β?transparent)␈α∞aspects␈α∞of␈α∞␈↓¬MYCIN␈↓'s␈α
behavior.␈α∞ If␈α∞the␈α∞facilities␈α
were
␈↓ α←␈↓␈↓ β?to be an effective debugging tool, they had to be accurate.

␈↓"β␈↓ α←␈↓␈↓ ββ(2)␈↓ β?␈↓↓Comprehensibility␈↓.␈α∪ Since␈α∀computer␈α∪consultants␈α∀are␈α∪intended
␈↓ α←␈↓␈↓ β?for␈α∃use␈α∃by␈α∀a␈α∃nonprogramming␈α∃audience,␈α∃the␈α∀explanations
␈↓ α←␈↓␈↓ β?generated␈α⊂by␈α⊂␈↓¬TEIRESIAS␈↓␈α⊂had␈α⊂to␈α⊂be␈α⊂tailored␈α⊂accordingly.␈α⊂ This
␈↓ α←␈↓␈↓ β?meant␈αrestrictions␈α
on␈αcontent␈αand␈α
vocabulary␈αand␈αan␈α
emphasis
␈↓ α←␈↓␈↓ β?on␈α↔brevity.␈α_ This␈α↔criterion␈α↔was␈α_the␈α↔main␈α↔source␈α_of␈α↔the
␈↓ α←␈↓␈↓ β?temptation␈α⊃to␈α⊃gloss␈α⊃over␈α⊃parts␈α⊃of␈α⊃␈↓¬MYCIN␈↓'s␈α⊃behavior,␈α⊃to␈α⊂avoid
␈↓ α←␈↓␈↓ β?having␈α⊃to␈α⊃justify␈α⊂in␈α⊃layman's␈α⊃terms␈α⊂the␈α⊃decisions␈α⊃that␈α⊂were
␈↓ α←␈↓␈↓ β?based on computational considerations.

␈↓"β␈↓ α←␈↓␈↓ ββ(3)␈↓ β?␈↓↓Human␈α_engineering␈↓.␈α_ Consideration␈α↔was␈α_also␈α_given␈α_to␈α↔a
␈↓ α←␈↓␈↓ β?collection␈αof␈α
user-oriented␈αfactors␈αlike␈α
ease␈αof␈αuse,␈α
power,␈αand
␈↓ α←␈↓␈↓ β?speed.

␈↓"β␈↓ α←␈↓␈↓ β?As␈α
might␈α
be␈αexpected,␈α
(1)␈α
and␈α
(2)␈αoccasionally␈α
conflict,␈α
in␈αpart␈α
because
␈↓ α←␈↓the␈αfacilities␈αare␈αpowerful␈αenough␈αto␈αallow␈αthe␈αuser␈αto␈αexamine␈αaspects␈αof␈αthe
␈↓ α←␈↓performance␈α∂program␈α∂not␈α∂normally␈α∞intended␈α∂for␈α∂display.␈α∂ As␈α∂one␈α∞example,
␈↓ α←␈↓the␈α
first␈αeight␈α
or␈αnine␈α
questions␈αof␈α
a␈α
consultation␈αare␈α
generated␈αas␈α
part␈αof␈α
the
␈↓ α←␈↓initialization␈α∃phase␈α∃of␈α∃the␈α∃program␈α∃and␈α∃hence␈α∃use␈α∃a␈α∃few␈α∃nonstandard
␈↓ α←␈↓mechanisms.␈α∞ Their␈α∞external␈α∂appearance␈α∞is␈α∞the␈α∂same␈α∞as␈α∞those␈α∂generated␈α∞by
␈↓ α←␈↓the␈α∀standard␈α∀method␈α∃of␈α∀backward␈α∀chaining␈α∀of␈α∃rules,␈α∀but␈α∀the␈α∃user␈α∀can
␈↓ α←␈↓(perhaps␈αunwittingly)␈αuncover␈αsome␈αperplexing␈αoperations␈αif␈αhe␈αexplores␈αthis
␈↓ α←␈↓part␈α∞of␈α∞the␈α∞process␈α∞with␈α
the␈α∞explanation␈α∞facilities.␈α∞ There␈α∞are␈α∞good␈α
reasons
␈↓ α←␈↓for␈α⊂all␈α∂of␈α⊂these␈α∂operations,␈α⊂but␈α∂it␈α⊂would␈α∂take␈α⊂some␈α∂extended␈α⊂discussion␈α∂to
␈↓ α←␈↓make them clear to a nonprogrammer.
␈↓ α←␈↓␈↓36    EXPLANATION␈↓ 
#3-3␈↓

␈↓"β␈↓ α←␈↓␈↓ β?Wherever␈α∩conflicts␈α∩did␈α∩arise,␈α⊃they␈α∩were␈α∩resolved␈α∩using␈α∩the␈α⊃design
␈↓ α←␈↓goals␈αin␈αthe␈αorder␈αlisted.␈α The␈αfacilities␈αpresent␈αan␈αaccurate␈αpicture␈αof␈αsystem
␈↓ α←␈↓performance;␈αthey␈αdo␈αso␈αas␈αcomprehensibly␈αas␈αpossible␈αand␈αattempt␈αto␈αbe␈α
fast
␈↓ α←␈↓and easy to use.

␈↓"β␈↓ α←␈↓␈↓α3-4    BASIC IDEAS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αbasic␈αideas␈α
behind␈αthe␈αdesign␈αof␈α
the␈αexplanation␈αfacilities␈αcan␈α
be
␈↓ α←␈↓viewed in terms of the four steps discussed below.

␈↓"β␈↓ α←␈↓␈↓ ββ(1)␈↓ β?␈↓↓Determine␈α⊗the␈α⊗program␈α⊗operation␈α⊗that␈α⊗is␈α⊗to␈α⊗be␈α⊗viewed␈α∃as
␈↓ α←␈↓↓␈↓ β?primitive.␈↓

␈↓"β␈↓ α←␈↓␈↓ β?This␈α⊗gives␈α⊗the␈α⊗smallest␈α⊗unit␈α∃of␈α⊗program␈α⊗behavior␈α⊗that␈α⊗can␈α∃be
␈↓ α←␈↓explained.␈α∀ Examples␈α∀further␈α∀on␈α∃will␈α∀demonstrate␈α∀that␈α∀it␈α∀is␈α∃possible␈α∀to
␈↓ α←␈↓generate␈α⊂different␈α∂degrees␈α⊂of␈α∂abstraction,␈α⊂but␈α∂the␈α⊂level␈α∂chosen␈α⊂in␈α⊂this␈α∂step
␈↓ α←␈↓determines␈α∞the␈α∞level␈α∞of␈α∞maximum␈α∞detail.␈α
 In␈α∞our␈α∞case,␈α∞the␈α∞invocation␈α∞of␈α
an
␈↓ α←␈↓individual rule was selected as the primitive operation.

␈↓"β␈↓ α←␈↓␈↓ ββ(2)␈↓ β?␈↓↓Augment␈αthe␈α
performance␈αprogram␈α
code␈αto␈α
leave␈αbehind␈αa␈α
record
␈↓ α←␈↓↓␈↓ β?of behavior at this level of detail.␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩result␈α⊃is␈α∩a␈α∩complete␈α⊃trace␈α∩of␈α∩program␈α⊃behavior,␈α∩a␈α∩history␈α⊃of
␈↓ α←␈↓performance.␈α∩The␈α∩relevant␈α∩sections␈α∩of␈α∩the␈α∩performance␈α∩program's␈α∩control
␈↓ α←␈↓structure␈α∞(i.e.,␈α∂the␈α∞inference␈α∞engine)␈α∂were␈α∞augmented␈α∞in␈α∂this␈α∞way␈α∞to␈α∂write␈α∞a
␈↓ α←␈↓history of rule invocation attempts.

␈↓"β␈↓ α←␈↓␈↓ ββ(3)␈↓ β?␈↓↓Select a global framework in which that trace can be understood.␈↓

␈↓"β␈↓ α←␈↓␈↓ β?This␈α
framework␈α
is␈α∞important␈α
especially␈α
where␈α∞computationally␈α
naive
␈↓ α←␈↓users␈α∪are␈α∪concerned.␈α∪ The␈α∪trace␈α∩provides␈α∪a␈α∪record␈α∪of␈α∪behavior,␈α∪but␈α∩the
␈↓ α←␈↓framework␈α∂supplies␈α∂a␈α∂way␈α∂of␈α∞understanding␈α∂that␈α∂behavior.␈α∂ Its␈α∂selection␈α∞is
␈↓ α←␈↓thus a central task in the construction of the facilities.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂framework␈α∞was␈α∂readily␈α∞supplied␈α∂by␈α∞the␈α∂performance␈α∞program's
␈↓ α←␈↓control␈αstructure--the␈αand/or␈α
goal␈αtree␈αoffered␈α
a␈αnatural␈αperspective␈αin␈α
which
␈↓ α←␈↓to␈α_view␈α_program␈α_behavior␈α_and␈α↔its␈α_explanation.␈α_ The␈α_tracing␈α_task␈α↔is
␈↓ α←␈↓structured␈α∀in␈α∃terms␈α∀of␈α∀tree␈α∃traversal,␈α∀and␈α∃the␈α∀user␈α∀is␈α∃offered␈α∀a␈α∃set␈α∀of
␈↓ α←␈↓commands␈α∂designed␈α∂around␈α∞it.␈α∂ The␈α∂goal-tree␈α∂view␈α∞is␈α∂of␈α∂course␈α∂specific␈α∞to
␈↓ α←␈↓this␈α∞particular␈α∂control␈α∞structure,␈α∞but␈α∂analogous␈α∞frameworks␈α∞are␈α∂possible␈α∞for
␈↓ α←␈↓other system designs.

␈↓"β␈↓ α←␈↓␈↓ ββ(4)␈↓ β?␈↓↓Design␈α≠a␈α≠program␈α~that␈α≠can␈α≠provide␈α~the␈α≠user␈α≠with␈α~an
␈↓ α←␈↓↓␈↓ β?interpretation of the trace.␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃program␈α∩should␈α⊃be␈α⊃capable␈α∩of␈α⊃systematic␈α⊃examination␈α∩of␈α⊃the
␈↓ α←␈↓trace␈α
and␈α∞should␈α
use␈α∞the␈α
framework␈α∞chosen␈α
to␈α∞provide␈α
an␈α∞interpretation␈α
of
␈↓ α←␈↓␈↓3-4␈↓ λzBASIC IDEAS    37␈↓

␈↓"β␈↓ α←␈↓events␈α⊃recorded␈α⊂there.␈α⊃ A␈α⊂body␈α⊃of␈α⊂code␈α⊃to␈α⊂do␈α⊃this␈α⊂forms␈α⊃the␈α⊂explanation
␈↓ α←␈↓program␈α∂in␈α∞␈↓¬TEIRESIAS␈↓␈α∂and␈α∞enables␈α∂the␈α∞user␈α∂to␈α∞examine␈α∂the␈α∞behavior␈α∂of␈α∞the
␈↓ α←␈↓performance␈αprogram.␈α
 It␈αinterprets␈αthe␈α
trace␈αin␈αterms␈α
of␈αthe␈αand/or␈α
goal-tree
␈↓ α←␈↓framework␈αand␈αprovides␈αa␈αset␈α
of␈αcommands␈αthat␈αallows␈αthe␈αuser␈α
to␈αexamine
␈↓ α←␈↓previous, current, or future (potential) behavior.

␈↓ α←␈↓These␈αideas␈αform␈α
the␈αfoundation␈αfor␈αthe␈α
basic␈αset␈αof␈αexplanation␈α
capabilities.
␈↓ α←␈↓They are illustrated below with a number of annotated examples.

␈↓"β␈↓ α←␈↓␈↓α3-5    TRACE OF SYSTEM PERFORMANCE:  EXPLANATIONS FOR
␈↓ α←␈↓α␈↓ β3PERFORMANCE VALIDATION␈↓
␈↓"β␈↓ α←␈↓␈↓ ¬GModern-day␈α≠computers␈α≠are␈α≠amazing␈α≠pieces␈α~of
␈↓ α←␈↓␈↓ ¬Gequipment,␈α~but␈α→most␈α~amazing␈α→of␈α~all␈α~are␈α→the
␈↓ α←␈↓␈↓ ¬Guncertain␈αgrounds␈αon␈αaccount␈αof␈αwhich␈α
we␈αattach
␈↓ α←␈↓␈↓ ¬Gany validity to their output.
␈↓"β␈↓ α←␈↓␈↓ 	5[Dijkstra72]
␈↓"β␈↓ α←␈↓␈↓ β?During␈α
a␈α∞consultation,␈α
the␈α
performance␈α∞program␈α
takes␈α∞the␈α
initiative,
␈↓ α←␈↓asking␈α
questions␈α∞of␈α
the␈α∞user.␈α
If␈α∞one␈α
of␈α∞those␈α
questions␈α∞seems␈α
inappropriate,
␈↓ α←␈↓the␈α∂user␈α∂can␈α∂interrupt␈α∂and␈α∂use␈α∂␈↓¬TEIRESIAS␈↓'s␈α∂explanation␈α∂facilities␈α∂to␈α∂find␈α∞out
␈↓ α←␈↓what's␈α∂going␈α∞on.␈α∂ By␈α∂examining␈α∞the␈α∂chain␈α∂of␈α∞reasoning␈α∂that␈α∂prompted␈α∞the
␈↓ α←␈↓question,␈α∀he␈α∀can␈α∀find␈α∀out␈α∀if␈α∀the␈α∀reasoning␈α∀was␈α∀motivated␈α∀by␈α∪plausible
␈↓ α←␈↓considerations.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈α
explanation␈α
is␈α
viewed␈α
in␈α
terms␈α
of␈α
traversal␈α
of␈α
the␈α∞and/or␈α
goal
␈↓ α←␈↓tree,␈αthe␈αcommand␈αset␈αis␈αdesigned␈αaccordingly.␈α The␈αtwo␈αbasic␈αcommands␈αare
␈↓ α←␈↓``why''␈α
and␈α
``how,''␈α
corresponding␈α
to␈α
ascent␈α
and␈α
descent␈α
of␈α
the␈α
tree.␈α
 It␈α
is␈α
the
␈↓ α←␈↓performance␈α⊂program's␈α⊂invocation␈α∂of␈α⊂rules␈α⊂in␈α∂a␈α⊂goal-directed␈α⊂fashion␈α∂that
␈↓ α←␈↓makes tree traversal operations plausible interpretations of the commands.
␈↓"β␈↓ α←␈↓␈↓ β?Several␈α
annotated␈α
examples␈α
are␈α
given␈α
below,␈α
in␈α
this␈α
and␈α
subsequent
␈↓ α←␈↓sections.␈α∩ In␈α∩each␈α∩case,␈α∩a␈α⊃single␈α∩question␈α∩from␈α∩the␈α∩consultation␈α∩has␈α⊃been
␈↓ α←␈↓extracted␈α∪and␈α∪the␈α∪explanation␈α∪facilities␈α∪have␈α∪been␈α∪used␈α∪to␈α∀examine␈α∪the
␈↓ α←␈↓reasoning␈α⊂behind␈α⊂it.␈α⊂ Fig.␈α⊂3-1␈α⊂shows␈α∂the␈α⊂part␈α⊂of␈α⊂the␈α⊂reasoning␈α⊂tree␈α∂being
␈↓ α←␈↓examined in the first example.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α∂the␈α∂trace␈α∂below,␈α∂computer␈α∂output␈α∂is␈α∂in␈α∂␈↓	mixed␈α∂upper␈α⊂and␈α∂lower
␈↓ α←␈↓	case␈α∞in␈α∂this␈α∞font␈↓,␈α∂user␈α∞responses␈α∞are␈α∂in␈α∞␈↓αBOLDFACE␈α∂CAPITALS␈↓,␈α∞while
␈↓ α←␈↓annotations are indented in ␈↓↓{italics inside braces}␈↓.

␈↓ α←␈↓	***************************************************************

␈↓ α←␈↓	In light of the site from which the culture was obtained and
␈↓ α←␈↓	the method of collection, do you feel that a significant number
␈↓ α←␈↓	of ORGANISM-1 were obtained?
␈↓ α←␈↓	** ␈↓αWHY␈↓	
␈↓ α←␈↓	[i.e., Why is it important to determine whether a ORGANISM-1
␈↓ α←␈↓	 was observed in significant numbers?]

␈↓ α←␈↓	1.0] This will aid in determining whether there is
␈↓ α←␈↓	     therapeutically significant disease associated with this
␈↓ α←␈↓	     occurrence of ORGANISM-1.
␈↓ α←␈↓␈↓38    EXPLANATION␈↓ 
#3-5␈↓

␈↓"β␈↓ α←␈↓	     It has already been established that
␈↓ α←␈↓	 [1.1] the method of collection is sterile,
␈↓ α←␈↓	     therefore, if:
␈↓ α←␈↓	 [1.2] the organism was observed in significant numbers, and
␈↓ α←␈↓	 [1.3] the site of the culture is not one of those which are
␈↓ α←␈↓	       normally sterile,
␈↓ α←␈↓	     then
␈↓ α←␈↓	 there is strongly suggestive evidence (.9) that there is
␈↓ α←␈↓	 therapeutically significant disease associated with this
␈↓ α←␈↓	 occurrence of the organism.
␈↓ α←␈↓	 [Also: there is strongly suggestive evidence (.8) that the
␈↓ α←␈↓	  organism is not a contaminant.]
␈↓ α←␈↓	 [RULE054]

␈↓ α←␈↓↓␈↓ βW{As␈α~Fig.␈α~3-1␈α≠shows,␈α~the␈α~issue␈α≠of␈α~``a␈α~significant␈α≠number␈α~of
␈↓ α←␈↓↓␈↓ βWorganisms''␈α∩is␈α∩one␈α⊃subgoal,␈α∩``therapeutically␈α∩significant␈α∩disease''␈α⊃is
␈↓ α←␈↓↓␈↓ βWanother␈α
at␈α
the␈α
next␈α
level␈α
up,␈αand␈α
rule␈α
54␈α
links␈α
them.␈α
 The␈α``reason''
␈↓ α←␈↓↓␈↓ βWfor␈α∃asking␈α∃the␈α⊗original␈α∃question,␈α∃then,␈α∃is␈α⊗to␈α∃invoke␈α∃rule␈α⊗54␈α∃to
␈↓ α←␈↓↓␈↓ βWdetermine␈α∞the␈α∞higher␈α∞subgoal,␈α∞and␈α∂the␈α∞answer␈α∞to␈α∞``why''␈α∞is␈α∂simply␈α∞a
␈↓ α←␈↓↓␈↓ βWdisplay of the goals and the rule linking them.
␈↓ α←␈↓↓␈↓ βWConsiderable␈α∀attention␈α∀has␈α∪been␈α∀paid␈α∀to␈α∪making␈α∀this␈α∀as␈α∀lucid␈α∪a
␈↓ α←␈↓↓␈↓ βWdisplay␈α as␈α possible.␈α  First,␈α since␈α ``why''␈α has␈α!several␈α other
␈↓ α←␈↓↓␈↓ βWinterpretations␈α∞and␈α∞no␈α∂more␈α∞extensive␈α∞natural␈α∂language␈α∞capabilities
␈↓ α←␈↓↓␈↓ βWare␈α⊂available,␈α⊂the␈α⊂question␈α⊂is␈α⊂expanded␈α⊂by␈α⊂␈↓¬TEIRESIAS␈↓↓␈α⊂to␈α⊂indicate␈α∂the
␈↓ α←␈↓↓␈↓ βWsense␈α⊂in␈α⊂which␈α⊂it␈α∂will␈α⊂be␈α⊂answered.␈α⊂ The␈α∂higher␈α⊂level␈α⊂goal␈α⊂is␈α∂then
␈↓ α←␈↓↓␈↓ βWpresented,␈αfollowed␈αby␈αthe␈αrule.␈α The␈αpremises␈αof␈αthe␈αrule␈αare␈αdivided
␈↓ α←␈↓↓␈↓ βWinto␈α∂those␈α∞already␈α∂established␈α∞and␈α∂those␈α∞yet␈α∂to␈α∞be␈α∂determined,␈α∞using
␈↓ α←␈↓↓␈↓ βWthe␈α
template␈α
mechanism␈α
described␈α
in␈α
chapter␈α
2.␈α
 Finally,␈α∞since␈α
rules
␈↓ α←␈↓↓␈↓ βWmay␈αhave␈αmultiple␈αconclusions␈αabout␈αdifferent␈αattributes,␈αthe␈αrelevant
␈↓ α←␈↓↓␈↓ βWconclusion is presented first and all others follow.}

␈↓ α←␈↓	** ␈↓αWHY␈↓	

␈↓ α←␈↓↓␈↓ βW{Further␈α_sections␈α_of␈α_the␈α→reasoning␈α_chain␈α_can␈α_be␈α→examined␈α_by
␈↓ α←␈↓↓␈↓ βWrepeating the command.}

␈↓ α←␈↓	[i.e., Why is it important to find out whether there is
␈↓ α←␈↓	 therapeutically significant disease associated with this
␈↓ α←␈↓	 occurrence of ORGANISM-1?]

␈↓ α←␈↓	[2.0]...in order to find out a locus of infection in PATIENT-1.
␈↓ α←␈↓	     It has already been established that:
␈↓ α←␈↓	 [2.1] the number of days since the specimen for the culture
␈↓ α←␈↓	       was obtained is less than 7,
␈↓ α←␈↓	     therefore, if:
␈↓ α←␈↓	 [2.2] there is therapeutically significant disease associated
␈↓ α←␈↓	       with this occurrence of the organism
␈↓ α←␈↓	     then
␈↓ α←␈↓	 it is definite (1.0) that the site of the culture is a locus
␈↓"β␈↓ α←␈↓␈↓3-5␈↓ ¬FEXPLANATIONS FOR PERFORMANCE VALIDATION    39␈↓

␈↓"β␈↓ α←␈↓	 of infection in the patient.
␈↓ α←␈↓	 [RULE021]

␈↓ α←␈↓	** ␈↓αWHY␈↓	
␈↓ α←␈↓	[i.e., Why is it important to determine a locus of infection
␈↓ α←␈↓	 in PATIENT-1?]

␈↓ α←␈↓	[3.0]...in order to find out about an organism or class of
␈↓ α←␈↓	     organisms for which therapy should cover.
␈↓ α←␈↓	     It has already been established that:
␈↓ α←␈↓	 [3.1] the site of the culture is blood, and
␈↓ α←␈↓	 [3.2] the stain of the organism is grampos, and
␈↓ α←␈↓	 [3.3] the portal of entry of the organism is GI,
␈↓ α←␈↓	     therefore, if:
␈↓ α←␈↓	 [3.4] [A - abdomen is a locus of infection in the patient, or
␈↓ α←␈↓	        B - pelvis is a locus of infection in the patient]
␈↓ α←␈↓	     then
␈↓ α←␈↓	 there is strongly suggestive evidence (.9) that
␈↓ α←␈↓	 enterobacteriaceae is a class of organisms for which therapy
␈↓ α←␈↓	 should cover.
␈↓ α←␈↓	 [RULE095]

␈↓ α←␈↓	** ␈↓αHOW [3.0]␈↓	

␈↓ α←␈↓↓␈↓ βW{HOW␈αcan␈αbe␈αused␈αto␈αclimb␈α``down''␈αthe␈αreasoning␈αchain␈αand␈αgives␈αa
␈↓ α←␈↓↓␈↓ βWpicture␈αof␈αall␈αthe␈α
ways␈αany␈αsubgoal␈αmay␈α
be␈αachieved.␈α In␈αthe␈α
examples
␈↓ α←␈↓↓␈↓ βWdemonstrating␈α∂its␈α∂use,␈α∂numbers␈α⊂following␈α∂the␈α∂command␈α∂refer␈α⊂to␈α∂the
␈↓ α←␈↓↓␈↓ βWcorresponding items above (e.g., the ␈↓	[3.0]␈↓↓ above).}

␈↓ α←␈↓	[i.e., How were you determining an organism or class of
␈↓ α←␈↓	 organisms for which therapy should cover?]

␈↓ α←␈↓	[4.0] At that point RULE095 was being used.  In addition,
␈↓ α←␈↓	 [4.1] RULE162 was used to conclude that bacteroides is an
␈↓ α←␈↓	       organism for which therapy should cover, and
␈↓ α←␈↓	 [4.2] RULE115 will be tried later, as well.

␈↓ α←␈↓↓␈↓ βW{Once␈α∀again␈α∃the␈α∀question␈α∀is␈α∃underspecified␈α∀and␈α∀is␈α∃expanded␈α∀by
␈↓ α←␈↓↓␈↓ βW␈↓¬TEIRESIAS␈↓↓.␈α_ Note␈α↔that␈α_the␈α_system␈α↔can␈α_indicate␈α_reasoning␈α↔chains
␈↓ α←␈↓↓␈↓ βWpreviously␈α
followed,␈α∞those␈α
currently␈α
being␈α∞examined,␈α
and␈α
those␈α∞yet␈α
to
␈↓ α←␈↓↓␈↓ βWbe tried.  Each can be further explored.}

␈↓ α←␈↓	** ␈↓αHOW [4.1]␈↓	
␈↓ α←␈↓	[i.e., How was RULE162 used?]

␈↓ α←␈↓	[5.0]  It has already been established that:
␈↓ α←␈↓	 [5.1] the site of the culture is blood, and
␈↓ α←␈↓	 [5.2] the portal of entry of the organism is GI, and
␈↓ α←␈↓	 [5.3] the patient is a compromised host.
␈↓ α←␈↓	       Therefore,
␈↓ α←␈↓	 it is definite (1.0) that bacteroides is an organism for
␈↓"β␈↓ α←␈↓␈↓40    EXPLANATION␈↓ 
#3-5␈↓

␈↓"β␈↓ α←␈↓	 which therapy should cover.
␈↓ α←␈↓	 [RULE162]

␈↓ α←␈↓	** ␈↓αHOW [4.2]␈↓	
␈↓ α←␈↓	[i.e. How will RULE115 be used?]

␈↓ α←␈↓	 If [6.0] brain-abscess is an infectious disease diagnosis
␈↓ α←␈↓	          for the patient,
␈↓ α←␈↓	 then     there is weakly suggestive evidence (.2) that
␈↓ α←␈↓	          streptococcus-anaerobic is an organism for which
␈↓ α←␈↓	          therapy should cover.
␈↓ α←␈↓	          [RULE115]
␈↓ α←␈↓␈↓3-5␈↓ ¬FEXPLANATIONS FOR PERFORMANCE VALIDATION    41␈↓





␈↓"␈↓ α←␈↓∧                  ⊂αααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧                  ~ what organisms ~
␈↓"␈↓ α←␈↓∧                  ~ should therapy ~
␈↓"␈↓ α←␈↓∧                  ~     cover      ~
␈↓"␈↓ α←␈↓∧                  %αααααααααααααααα$
␈↓"␈↓ α←␈↓∧                          ↑
␈↓"␈↓ α←␈↓∧        RULE162           ~ RULE095          RULE115
␈↓"␈↓ α←␈↓∧                          ~



␈↓"␈↓ α←␈↓∧⊂αααααα⊃  ⊂αααααααα⊃  ⊂αααααααα⊃  ⊂ααααααα⊃  ⊂ααααααα⊃
␈↓"␈↓ α←␈↓∧~site =~  ~gram =  ~  ~portal =~  ~locus =~  ~locus =~
␈↓"␈↓ α←␈↓∧~blood ~  ~positive~  ~GI tract~  ~abdomen~  ~pelvis ~
␈↓"␈↓ α←␈↓∧%αααααα$  %αααααααα$  %αααααααα$  %ααααααα$  %ααααααα$
␈↓"␈↓ α←␈↓∧                                            ↑
␈↓"␈↓ α←␈↓∧                                            ~ RULE021
␈↓"␈↓ α←␈↓∧                                            ~


␈↓"␈↓ α←␈↓∧                         ⊂ααααααααααααα⊃        ⊂ααααααααααααα⊃
␈↓"␈↓ α←␈↓∧                         ~  num. days  ~        ~ significant ~
␈↓"␈↓ α←␈↓∧                         ~ less than 7 ~        ~   disease   ~
␈↓"␈↓ α←␈↓∧                         %ααααααααααααα$        %ααααααααααααα$
␈↓"␈↓ α←␈↓∧                                                      ↑
␈↓"␈↓ α←␈↓∧                                                      ~ RULE054
␈↓"␈↓ α←␈↓∧                                                      ~


␈↓"␈↓ α←␈↓∧              ⊂ααααααααααααα⊃  ⊂ααααααααααααα⊃  ⊂ααααααααααααα⊃
␈↓"␈↓ α←␈↓∧              ~  collection ~  ~ significant ~  ~  nonsterile ~
␈↓"␈↓ α←␈↓∧              ~   sterile   ~  ~   number    ~  ~    site     ~
␈↓"␈↓ α←␈↓∧              %ααααααααααααα$  %ααααααααααααα$  %ααααααααααααα$

␈↓"␈↓ α←␈↓∧)  ␈↓-- conjunction␈↓	
␈↓"␈↓ α←␈↓	)) ␈↓-- disjunction


␈↓"␈↓ α←␈↓α␈↓ αzFig. 3-1.    Reasoning tree for the first set of explanation examples.    
␈↓ α←␈↓␈↓42    EXPLANATION␈↓ 
#3-6␈↓

␈↓"β␈↓ α←␈↓␈↓α3-6    THE NEED FOR AN INFORMATION METRIC␈↓
␈↓"β␈↓ α←␈↓␈↓ ¬G[In␈α∞an␈α∞explanation]␈α∂we␈α∞must␈α∞not␈α∂carry␈α∞reasoning
␈↓ α←␈↓␈↓ ¬Gtoo␈α∂far␈α∞back,␈α∂or␈α∞the␈α∂length␈α∞of␈α∂our␈α∂argument␈α∞will
␈↓ α←␈↓␈↓ ¬Gcause␈α⊃obscurity;␈α⊂neither␈α⊃must␈α⊃we␈α⊂put␈α⊃in␈α⊃all␈α⊂the
␈↓ α←␈↓␈↓ ¬Gsteps␈α	that␈α
lead␈α	to␈α
our␈α	conclusion,␈α
or␈α	we␈α
shall␈α	waste
␈↓ α←␈↓␈↓ ¬Gwords in saying what is manifest.
␈↓"β␈↓ α←␈↓␈↓ 	.[Aristotle26]
␈↓"β␈↓ α←␈↓␈↓ β?One␈αproblem␈αanticipated␈αin␈αthe␈αuse␈αof␈αthe␈αWHY␈αcommand,␈α
and␈αone
␈↓ α←␈↓that␈α⊂is␈α⊂common␈α∂to␈α⊂explanations␈α⊂in␈α∂general,␈α⊂is␈α⊂the␈α∂issue␈α⊂of␈α⊂an␈α∂appropriate
␈↓ α←␈↓level of sophistication and detail.  It is generally of little use to discover that

␈↓"β␈↓ α←␈↓	If:  1) The gram stain of the organism is grampos, and
␈↓"β␈↓ α←␈↓	     2) The morphology of the organism is rod
␈↓"β␈↓ α←␈↓	Then:  It is definite (1.0) that the category of the organism
␈↓"β␈↓ α←␈↓	       is grampos-rods.

␈↓"β␈↓ α←␈↓α␈↓ ¬6Fig. 3-2.    RULE140.    

␈↓ α←␈↓Depending␈α
on␈α
the␈α
individual␈α
user,␈α
it␈α
might␈αbe␈α
best␈α
to␈α
display␈α
all␈α
steps␈α
in␈αa
␈↓ α←␈↓reasoning␈αchain,␈α
to␈αomit␈αthose␈α
that␈αare␈α
definitional␈αor␈αtrivial,␈α
or,␈αfor␈αthe␈α
most
␈↓ α←␈↓sophisticated␈α⊃user,␈α⊂to␈α⊃display␈α⊂only␈α⊃the␈α⊂highlights.␈α⊃ This␈α⊂presumes␈α⊃that␈α⊂we
␈↓ α←␈↓have␈αsome␈αidea␈αof␈αwhat␈αconstitutes␈α``the␈αdetails.'' ␈αIn␈αterms␈αof␈αthe␈αgoal␈αtree,␈αit
␈↓ α←␈↓means␈αknowing␈αhow␈α``far''␈αit␈αis␈αconceptually␈αfrom␈αone␈αnode␈αto␈αanother,␈αwhich
␈↓ α←␈↓is␈αdifficult␈α
since␈αthis␈α
depends␈αvery␈α
much␈αon␈α
the␈α(constantly␈α
changing)␈αstate␈α
of
␈↓ α←␈↓the␈α
user's␈α
knowledge.␈α
 It␈α
is␈α
also␈α
very␈α
important␈α
since␈α
the␈α
best␈α
explanations␈α
are
␈↓ α←␈↓those␈α≠based␈α≠on␈α≠a␈α≠clear␈α≠understanding␈α≠of␈α≠the␈α≠extent␈α≠of␈α≠the␈α~user's
␈↓ α←␈↓comprehension.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α∞a␈α∞very␈α
rough␈α∞analogy␈α∞to␈α
information␈α∞theory,␈α∞we␈α
use␈α∞-(log␈α∞CF)␈α
to
␈↓ α←␈↓provide␈α∪this␈α∪measure.␈α∪ Definitional␈α∪rules␈α∩(with␈α∪CF␈α∪=␈α∪1.0)␈α∪thus␈α∪have␈α∩no
␈↓ α←␈↓information,␈αwhile␈α
those␈αwith␈α
smaller␈αCFs␈αhave␈α
progressively␈αmore.␈α
 This␈αis
␈↓ α←␈↓clearly␈α∀imperfect.␈α∀ It␈α∀does␈α∀not␈α∀take␈α∀into␈α∀account␈α∀the␈α∀state␈α∀of␈α∀the␈α∀user's
␈↓ α←␈↓knowledge,␈α~and␈α~since␈α~CFs␈α~are␈α~not␈α~probabilities␈α~there␈α~is␈α≠no␈α~formal
␈↓ α←␈↓justification␈α∩that␈α⊃-(log␈α∩CF)␈α⊃is␈α∩a␈α⊃meaningful␈α∩number.␈α⊃ It's␈α∩primary␈α⊃utility,
␈↓ α←␈↓however,␈αis␈α
as␈αa␈α
``dial''␈αwith␈α
which␈αthe␈α
user␈αcan␈α
adjust␈αthe␈α
level␈αof␈α
detail␈αin
␈↓ α←␈↓the␈α⊃explanations.␈α⊃ Absolute␈α⊂settings␈α⊃are␈α⊃less␈α⊂important␈α⊃than␈α⊃the␈α⊃ability␈α⊂to
␈↓ α←␈↓make relative adjustments.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩adjustment␈α∪is␈α∩made␈α∪via␈α∩an␈α∩optional␈α∪argument␈α∩to␈α∪the␈α∩WHY
␈↓ α←␈↓command.␈α
 The␈αentire␈α
``distance''␈αfrom␈α
the␈αcurrent␈α
node␈αto␈α
the␈αtop␈α
of␈αthe␈α
goal
␈↓ α←␈↓tree␈α
is␈α∞normalized␈α
to␈α
10,␈α∞and␈α
the␈α∞argument␈α
following␈α
the␈α∞WHY␈α
is␈α∞taken␈α
as
␈↓ α←␈↓some␈α
fraction␈α∞of␈α
that␈α∞distance.␈α
 Thus,␈α∞rather␈α
than␈α∞the␈α
multiple␈α∞WHYs␈α
seen
␈↓ α←␈↓above, the user might have asked ``WHY 4'':

␈↓"β␈↓ α←␈↓	In light of the site from which the culture was obtained, and
␈↓"β␈↓ α←␈↓	the method of collection, do you feel that a significant number
␈↓"β␈↓ α←␈↓	of ORGANISM-1 were obtained?
␈↓"β␈↓ α←␈↓	** ␈↓αWHY 4␈↓	
␈↓"β␈↓ α←␈↓	[i.e., Why is it important to determine whether a significant
␈↓"β␈↓ α←␈↓	number of ORGANISM-1 were obtained?]
␈↓ α←␈↓␈↓3-6␈↓ ε~THE NEED FOR AN INFORMATION METRIC    43␈↓

␈↓"β␈↓ α←␈↓	We are trying to find out whether the organism has been
␈↓"β␈↓ α←␈↓	observed in significant numbers, in order to determine an
␈↓"β␈↓ α←␈↓	organism or class of organisms for which therapy should cover.

␈↓"β␈↓ α←␈↓α␈↓ ∧JFig. 3-3.    A `high-level' explanation.    

␈↓ α←␈↓Since␈αthe␈αthree␈αsteps␈αin␈αthe␈α
chain␈αshown␈αin␈αFig.␈α3-1␈αconstitute␈α
roughly␈αtwo-
␈↓ α←␈↓fifths␈α
of␈α
the␈α``distance''␈α
to␈α
the␈αtop,␈α
they␈α
are␈αcompressed␈α
into␈α
a␈α
single␈αanswer.
␈↓ α←␈↓This␈α∞command␈α∞may␈α∞be␈α
repeated␈α∞as␈α∞often␈α∞as␈α
necessary,␈α∞allowing␈α∞the␈α∞user␈α
to
␈↓ α←␈↓follow the reasoning chain in step sizes of his own choosing.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α↔user␈α⊗may␈α↔occasionally␈α↔choose␈α⊗a␈α↔step␈α↔size␈α⊗that␈α↔is␈α↔too␈α⊗big,
␈↓ α←␈↓compressing␈α∀too␈α∪many␈α∀steps␈α∪into␈α∀a␈α∪single␈α∀answer,␈α∪leaving␈α∀him␈α∀with␈α∪an
␈↓ α←␈↓incomprehensible␈α⊂explanation.␈α⊂ In␈α⊂this␈α⊂case␈α⊂he␈α⊂can␈α⊂invoke␈α⊂the␈α∂EXPLAIN
␈↓ α←␈↓command,␈α∪which␈α∩will␈α∪cover␈α∩the␈α∪same␈α∩starting␈α∪and␈α∩ending␈α∪points␈α∪in␈α∩the
␈↓ α←␈↓reasoning,␈α∂but␈α∂in␈α∞more␈α∂detail.␈α∂ Thus,␈α∞if␈α∂the␈α∂explanation␈α∞in␈α∂Fig.␈α∂3-3␈α∞above
␈↓ α←␈↓proved␈α∀to␈α∀be␈α∀too␈α∀obscure,␈α∀the␈α∀user␈α∀might␈α∀say␈α∀EXPLAIN␈α∀and␈α∀have␈α∪it
␈↓ α←␈↓expanded␈αout␈αin␈αcomplete␈αdetail.␈α He␈αmight␈αalso␈αsay␈αEXPLAIN␈α2,␈αto␈αhave␈αit
␈↓ α←␈↓expanded␈α
out␈αin␈α
steps␈α
roughly␈αhalf␈α
as␈α
big␈αas␈α
in␈α
Fig.␈α3-3␈α
(the␈α
argument␈αhas
␈↓ α←␈↓the same meaning for both the WHY and EXPLAIN commands).

␈↓"β␈↓ α←␈↓␈↓α3-7    MORE SOPHISTICATED HOWS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Previous␈αexamples␈αhave␈αdemonstrated␈α
that␈αHOW␈αquestions␈αcan␈α
have
␈↓ α←␈↓a␈α⊂range␈α⊃of␈α⊂possible␈α⊂meanings,␈α⊃depending␈α⊂on␈α⊂the␈α⊃referent␈α⊂of␈α⊃the␈α⊂question.
␈↓ α←␈↓Examples␈αabove␈αshowed␈α
its␈αuse␈αin␈α
determining␈αhow␈αthe␈α
value␈αof␈αan␈α
attribute
␈↓ α←␈↓was␈α
reached␈αand␈α
how␈αdifferent␈α
rules␈αhad␈α
been␈α(or␈α
might␈αbe)␈α
used.␈α
 A␈αmore
␈↓ α←␈↓complex␈αsituation␈α
arises␈αif␈α
the␈αquestion␈αrefers␈α
to␈αa␈α
clause␈αfrom␈α
a␈αpremise␈αof␈α
a
␈↓ α←␈↓rule, as in:

␈↓"β␈↓ α←␈↓	Is the patient's illness with ORGANISM-1 a hospital-acquired
␈↓"β␈↓ α←␈↓	infection?
␈↓"β␈↓ α←␈↓	** ␈↓αWHY␈↓	
␈↓"β␈↓ α←␈↓	[i.e. WHY is it important to determine whether or not the
␈↓"β␈↓ α←␈↓	 infection with ORGANISM-1 was acquired while the patient was
␈↓"β␈↓ α←␈↓	 hospitalized?]

␈↓"β␈↓ α←␈↓	[2.0] This will aid in determining the category of ORGANISM-1.
␈↓"β␈↓ α←␈↓	      It has already been established that
␈↓"β␈↓ α←␈↓	 [2.1] the gram stain of ORGANISM-1 is gramneg, and
␈↓"β␈↓ α←␈↓	 [2.2] the morphology of ORGANISM-1 is rod, and
␈↓"β␈↓ α←␈↓	 [2.3] the aerobicity of ORGANISM-1 is facultative
␈↓"β␈↓ α←␈↓	      Therefore, if
␈↓"β␈↓ α←␈↓	 [2.4] the infection with ORGANISM-1 was not acquired while
␈↓"β␈↓ α←␈↓	       the patient was hospitalized
␈↓"β␈↓ α←␈↓	      then
␈↓"β␈↓ α←␈↓	 there is strongly suggestive evidence (.8) that the category
␈↓"β␈↓ α←␈↓	 of ORGANISM-1 is enterobacteriaceae
␈↓"β␈↓ α←␈↓	 [ALSO: there is weakly suggestive evidence (.1) that the
␈↓"β␈↓ α←␈↓	  identity of ORGANISM-1 is pseudomonas-aeruginosa]
␈↓"β␈↓ α←␈↓	 [RULE037]
␈↓"β␈↓ α←␈↓␈↓44    EXPLANATION␈↓ 
#3-7␈↓

␈↓"β␈↓ α←␈↓	** ␈↓αHOW [2.3]␈↓	
␈↓"β␈↓ α←␈↓	[i.e. HOW was it established that the aerobicity of ORGANISM-1
␈↓"β␈↓ α←␈↓	 is facultative?]

␈↓"β␈↓ α←␈↓	The following were used:
␈↓"β␈↓ α←␈↓	 [3.1] RULE027 indicated there is strongly suggestive evidence
␈↓"β␈↓ α←␈↓	       (.8) that the aerobicity of ORGANISM-1 is facultative
␈↓"β␈↓ α←␈↓	 [3.2] RULE033 indicated there is weakly suggestive evidence
␈↓"β␈↓ α←␈↓	       (.2) that the aerobicity of ORGANISM-1 is not
␈↓"β␈↓ α←␈↓	       facultative.

␈↓"β␈↓ α←␈↓	Since this gave a cumulative CF of (.6) for facultative, the
␈↓"β␈↓ α←␈↓	evidence indicates that the aerobicity of ORGANISM-1 is
␈↓"β␈↓ α←␈↓	facultative.

␈↓"β␈↓ α←␈↓␈↓ β?To␈α~answer␈α≠the␈α~question␈α~about␈α≠aerobicity␈α~accurately,␈α≠we␈α~must
␈↓ α←␈↓determine␈α_from␈α_the␈α_execution␈α→trace␈α_the␈α_subset␈α_of␈α_rules␈α→that␈α_actually
␈↓ α←␈↓contributed␈α⊃to␈α⊃the␈α⊃conclusion␈α⊃mentioned␈α⊃in␈α⊂clause␈α⊃[2.3].␈α⊃In␈α⊃this␈α⊃case␈α⊃it␈α⊂is
␈↓ α←␈↓straightforward,␈α↔since␈α⊗the␈α↔set␈α↔of␈α⊗all␈α↔rules␈α⊗invoked␈α↔about␈α↔aerobicity␈α⊗is
␈↓ α←␈↓available␈α⊂from␈α⊂the␈α⊂execution␈α⊃trace␈α⊂and␈α⊂the␈α⊂appropriate␈α⊂subset␈α⊃(those␈α⊂that
␈↓ α←␈↓concluded␈α∃that␈α∃aerobicity␈α∀was␈α∃facultative)␈α∃can␈α∀be␈α∃determined␈α∃using␈α∀the
␈↓ α←␈↓template mechanism to examine the conclusion of each of them.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∃second␈α∃example,␈α∀from␈α∃a␈α∃different␈α∀consultation,␈α∃shows␈α∃a␈α∀more
␈↓ α←␈↓complex response.

␈↓"β␈↓ α←␈↓	Is PATIENT-1 immunosuppressed or immunodeficient?
␈↓"β␈↓ α←␈↓	** ␈↓αWHY␈↓	
␈↓"β␈↓ α←␈↓	[i.e. Why is it important to determine if the patient is
␈↓"β␈↓ α←␈↓	 immunosuppressed or immunodeficient?]

␈↓"β␈↓ α←␈↓	[5.0] This will aid in determining the category of ORGANISM-1.
␈↓"β␈↓ α←␈↓	      It has already been established that
␈↓"β␈↓ α←␈↓	 [5.1] the site of CULTURE-1 is blood, and
␈↓"β␈↓ α←␈↓	 [5.2] the gram stain of ORGANISM-1 is gramneg, and
␈↓"β␈↓ α←␈↓	 [5.3] the morphology of ORGANISM-1 is rod,
␈↓"β␈↓ α←␈↓	      Therefore, if
␈↓"β␈↓ α←␈↓	 [5.4] the patient is immunosuppressed or immunodeficient,
␈↓"β␈↓ α←␈↓	      then
␈↓"β␈↓ α←␈↓	 there is weakly suggestive evidence (.3) that the category of
␈↓"β␈↓ α←␈↓	 ORGANISM-1 is enterobacteriaceae
␈↓"β␈↓ α←␈↓	 [RULE085]

␈↓"β␈↓ α←␈↓	**␈↓αWHY␈↓	
␈↓"β␈↓ α←␈↓	[i.e. Why is it important to determine the category of
␈↓"β␈↓ α←␈↓	 ORGANISM-1?]

␈↓"β␈↓ α←␈↓	[6.0] This will aid in determining the identity of ORGANISM-1.
␈↓"β␈↓ α←␈↓	      It has already been established that
␈↓"β␈↓ α←␈↓	 [6.1] the site of CULTURE-1 is blood, and
␈↓"β␈↓ α←␈↓	 [6.2] the gram stain of ORGANISM-1 is gramneg, and
␈↓"β␈↓ α←␈↓	 [6.3] the morphology of ORGANISM-1 is rod,
␈↓"β␈↓ α←␈↓	      Therefore, if
␈↓ α←␈↓␈↓3-7␈↓ π2MORE SOPHISTICATED HOWS    45␈↓

␈↓"β␈↓ α←␈↓	 [6.4] the category of ORGANISM-1 is not known with certainty,
␈↓"β␈↓ α←␈↓	      then
␈↓"β␈↓ α←␈↓	 there is weakly suggestive evidence (.3) that the identity of
␈↓"β␈↓ α←␈↓	 ORGANISM-1 is proteus.
␈↓"β␈↓ α←␈↓	 [RULE093]

␈↓"β␈↓ α←␈↓	** ␈↓αHOW [6.4]␈↓	
␈↓"β␈↓ α←␈↓	[i.e. How will it be established that the category of
␈↓"β␈↓ α←␈↓	 ORGANISM-1 is known, but not with certainty?]

␈↓"β␈↓ α←␈↓	[7.0] Currently RULE085 is being used.
␈↓"β␈↓ α←␈↓	      In addition, the following was also used:
␈↓"β␈↓ α←␈↓	 [7.1] RULE084 indicated that there is strongly suggestive
␈↓"β␈↓ α←␈↓	       evidence (.8) that the category of the organism is
␈↓"β␈↓ α←␈↓	       enterobacteriaceae.
␈↓"β␈↓ α←␈↓	      Based on current patient data, the following may also
␈↓"β␈↓ α←␈↓	      prove useful later in the consultation:
␈↓"β␈↓ α←␈↓	 [7.2] RULE176
␈↓"β␈↓ α←␈↓	 [7.3] RULE021

␈↓"β␈↓ α←␈↓	If, after trying these rules, the cumulative CF of any value of
␈↓"β␈↓ α←␈↓	the category of ORGANISM-1 is equal to or greater than .2 and
␈↓"β␈↓ α←␈↓	less than 1.0, it will be established that the category of
␈↓"β␈↓ α←␈↓	ORGANISM-1 is known, but not with certainty.

␈↓"β␈↓ α←␈↓␈↓ β?In␈α≤this␈α≤case␈α≠the␈α≤performance␈α≤program␈α≠has␈α≤not␈α≤yet␈α≠finished
␈↓ α←␈↓determining␈αthe␈αorganism's␈αidentity,␈αand␈αthe␈αuser␈αhas␈αasked␈αhow␈αthe␈αprocess
␈↓ α←␈↓will␈α∪establish␈α∪the␈α∪answer␈α∪with␈α∩the␈α∪indicated␈α∪range␈α∪of␈α∪certainty.␈α∩ Several
␈↓ α←␈↓sources␈α
of␈α
information␈αare␈α
required␈α
to␈αprovide␈α
an␈α
accurate␈α
and␈αinformative
␈↓ α←␈↓response.␈α∩ First,␈α∩␈↓¬TEIRESIAS␈↓␈α∪needs␈α∩some␈α∩indication␈α∪of␈α∩the␈α∩``meaning''␈α∪of␈α∩the
␈↓ α←␈↓predicate function ``known, but not with certainty.''
␈↓"β␈↓ α←␈↓␈↓ β?There␈α
is␈α
a␈α
simple␈α
framework␈αin␈α
which␈α
all␈α
of␈α
the␈α
predicate␈αfunctions
␈↓ α←␈↓can␈α
be␈α∞viewed,␈α
and␈α
which␈α∞provides␈α
the␈α
basis␈α∞for␈α
our␈α
representation␈α∞of␈α
this
␈↓ α←␈↓information.␈α Each␈αfunction␈αcan␈αbe␈αseen␈αas␈αsome␈αsegment␈αof␈αthe␈αnumber␈αline
␈↓ α←␈↓[-1.0,␈α∂1.0]␈α∂that␈α∂defines␈α∂the␈α∂range␈α∂of␈α∂certainty␈α∂factors␈α∂(several␈α∂examples␈α∂are
␈↓ α←␈↓indicated␈α≤in␈α≠Fig.␈α≤3-4␈α≠below,␈α≤see␈α≠[Shortliffe76]␈α≤for␈α≠the␈α≤entire␈α≠list).
␈↓ α←␈↓Associated␈α⊂with␈α∂each␈α⊂predicate␈α∂function␈α⊂is␈α∂a␈α⊂numerical␈α∂definition␈α⊂in␈α∂these
␈↓ α←␈↓terms.
␈↓ α←␈↓␈↓46    EXPLANATION␈↓ 
#3-7␈↓


␈↓"␈↓ α←␈↓∧          [  ←α    NOTSAME    α→  ]

␈↓"␈↓ α←␈↓∧                                  ( ←α   SAME  α→ ]

␈↓"␈↓ α←␈↓∧                                  [← NOTDEFINITE →)

␈↓"␈↓ α←␈↓∧          εαααααααααααααααβαααβαααβαααααααααααααααλ

␈↓"␈↓ α←␈↓∧        -1.0            -.2   0  .2              1.0


␈↓"␈↓ α←␈↓∧[ -- closed interval
␈↓"␈↓ α←␈↓∧( -- open interval

␈↓"β␈↓ α←␈↓α␈↓ β≠Fig. 3-4.    Predicate functions: Segments of the number line.    

␈↓ α←␈↓For␈α␈↓	NOTDEFINITE␈↓␈α(the␈αpredicate␈αfunction␈αused␈αin␈αclause␈α[6.4]),␈αthe␈αdefinition
␈↓ α←␈↓is

␈↓"β␈↓ α←␈↓	␈↓ ∧o(AND (GTE CF .2) (LT CF 1.0))

␈↓ α←␈↓Since␈α⊃experience␈α⊃has␈α⊃shown␈α⊃that␈α⊃the␈α⊃``canned''␈α⊃English␈α⊃phrases␈α⊂associated
␈↓ α←␈↓with␈α∂the␈α∂predicate␈α∂functions␈α∂are␈α∞often␈α∂unclear,␈α∂␈↓¬TEIRESIAS␈↓␈α∂prints␈α∂the␈α∞interval
␈↓ α←␈↓definition␈α∞at␈α∞the␈α∞end␈α∞of␈α∞the␈α
explanation.␈α∞ This␈α∞makes␈α∞clear␈α∞to␈α∞the␈α∞user␈α
the
␈↓ α←␈↓criterion␈α
used␈α
to␈α∞judge␈α
the␈α
truth␈α∞or␈α
falsity␈α
of␈α
a␈α∞clause.␈α
 In␈α
the␈α∞current␈α
case,
␈↓ α←␈↓for␈α
example,␈αit␈α
is␈αjudged␈α
to␈αbe␈α
true␈αthat␈α
``␈↓	the␈αcategory␈α
of␈α
ORGANISM-1␈αis
␈↓ α←␈↓	not␈αknown␈α
with␈αcertainty␈↓''␈α
if␈αsome␈αvalue␈α
has␈αbeen␈α
established␈αfor␈α
it␈αwith
␈↓ α←␈↓a CF greater than or equal to .2 and less than 1.0.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
set␈α
of␈α
answers␈α
obtained␈αby␈α
the␈α
user␈α
during␈α
the␈α
consultation␈αis␈α
the
␈↓ α←␈↓second␈α⊗major␈α⊗source␈α⊗of␈α⊗information.␈α∃ This␈α⊗set␈α⊗makes␈α⊗possible␈α⊗a␈α∃well-
␈↓ α←␈↓constrained␈α≥list␈α≥of␈α≥rules␈α≡that␈α≥``␈↓	may␈α≥prove␈α≥useful␈α≥later␈α≡in␈α≥the
␈↓ α←␈↓	consultation␈↓''␈α
(as␈α
in␈α
[7.2]␈α
and␈α
[7.3]␈α
above).␈α
 The␈α
entire␈α
list␈α
of␈α
rules␈α
about
␈↓ α←␈↓identity␈α
is␈α
available␈α
internally;␈α
the␈α
subset␈α
of␈α
them␈α
yet␈α
to␈α
be␈α
invoked␈α∞can␈α
be
␈↓ α←␈↓determined␈αby␈α
reference␈αto␈α
the␈αexecution␈αtrace.␈α
 The␈αpremise␈α
of␈αeach␈αof␈α
these
␈↓ α←␈↓is␈α∂checked␈α⊂(using␈α∂the␈α⊂partial␈α∂evaluation␈α∂technique␈α⊂discussed␈α∂in␈α⊂chapter␈α∂2),
␈↓ α←␈↓disqualifying␈α
those␈α∞that␈α
are␈α∞known␈α
to␈α∞be␈α
false␈α∞based␈α
on␈α∞currently␈α
available
␈↓ α←␈↓information.␈α∞ This␈α∞helps␈α
␈↓¬TEIRESIAS␈↓␈α∞keep␈α∞the␈α
answer␈α∞brief␈α∞and␈α∞avoid␈α
leading
␈↓ α←␈↓the user down a path in the reasoning tree known to be a dead end.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αmajor␈αcontribution␈α
of␈αthese␈αvarious␈α
sources␈αof␈αinformation␈α
is␈αto
␈↓ α←␈↓allow␈α
␈↓¬TEIRESIAS␈↓␈αto␈α
construct␈αexplanations␈α
that␈αare␈α
as␈αsharply␈α
focused,␈αprecise,
␈↓ α←␈↓and␈αinformative␈αas␈αpossible.␈α The␈αsystem's␈α
ability␈αto␈αconstrain␈αthe␈αset␈αof␈α
rules
␈↓ α←␈↓considered␈α∞means␈α∞that␈α∞the␈α∞user␈α∞will␈α∞not␈α∞be␈α∞inundated␈α∞by␈α∞large␈α∂numbers␈α∞of
␈↓ α←␈↓possible␈αpaths,␈αeven␈αwhen␈αthere␈αare␈αquite␈αa␈αfew␈αrules␈αin␈αthe␈αknowledge␈αbase.
␈↓ α←␈↓The␈α∪system's␈α∪ability␈α∪to␈α∪present␈α∪some␈α∪indication␈α∪of␈α∪the␈α∪definition␈α∪of␈α∩the
␈↓ α←␈↓predicate␈α
functions␈α
means␈α
that␈α
it␈α
can␈α
be␈α
terse␈α
in␈α
its␈α
normal␈α
translation␈α
of␈αa
␈↓ α←␈↓␈↓3-7␈↓ π2MORE SOPHISTICATED HOWS    47␈↓

␈↓"β␈↓ α←␈↓rule␈αand␈α
still␈αgive␈αthe␈α
user␈αsome␈α
idea␈αof␈αthe␈α
less␈αobvious␈α
aspects␈αof␈αthe␈α
model
␈↓ α←␈↓of confirmation.

␈↓"β␈↓ α←␈↓␈↓α3-8    SPECIAL PURPOSE RESPONSES␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃goal-tree␈α⊃framework␈α⊃accounts␈α⊃for␈α⊃a␈α⊃very␈α⊃large␈α⊃part␈α⊃of␈α⊂␈↓¬MYCIN␈↓'s
␈↓ α←␈↓performance,␈α
but␈α
not␈α
all␈α
of␈α
it.␈α
 In␈α
order␈α
to␈α
meet␈α
the␈α
primary␈α
design␈α
goal␈α
of
␈↓ α←␈↓accuracy,␈α∞it␈α∂was␈α∞necessary␈α∞to␈α∂implement␈α∞a␈α∞small␈α∂number␈α∞of␈α∂special␈α∞purpose
␈↓ α←␈↓responses in ␈↓¬TEIRESIAS␈↓'s explanation program.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αdescription␈αof␈αthe␈αcontrol␈α
structure␈αin␈αchapter␈α2␈αnoted␈α
that␈αeach
␈↓ α←␈↓object␈α∪(a␈α∩patient,␈α∪culture,␈α∩organism,␈α∪etc.)␈α∩has␈α∪a␈α∩number␈α∪of␈α∪its␈α∩associated
␈↓ α←␈↓attributes␈α⊂marked␈α⊂as␈α⊂``primary''␈α⊂and␈α⊂that␈α⊂these␈α⊂attributes␈α⊂are␈α⊂automatically
␈↓ α←␈↓traced␈α
when␈α
the␈α
object␈α
is␈α
instantiated.␈↓
2␈↓␈α
Since␈α
this␈α
process␈α
is␈α
not␈α
part␈α
of␈α
the
␈↓ α←␈↓standard␈α⊂goal-tree␈α⊂search,␈α⊂a␈α⊃separate␈α⊂mechanism␈α⊂is␈α⊂required␈α⊂to␈α⊃provide␈α⊂a
␈↓ α←␈↓response␈αto␈αa␈αWHY␈αinquiry␈αat␈αthis␈αpoint.␈α An␈αexample␈αis␈αshown␈αbelow: ␈αIt␈α
is
␈↓ α←␈↓a brief statement of fact, augmented by a very simple form of generalization.

␈↓"β␈↓ α←␈↓	From what site was the specimen for CULTURE-1 taken?
␈↓"β␈↓ α←␈↓	** ␈↓αWHY␈↓	

␈↓"β␈↓ α←␈↓	[i.e. Why is it important to determine the site of CULTURE-1?]
␈↓"β␈↓ α←␈↓	[8.0] There are 2 standard attributes that are important to
␈↓"β␈↓ α←␈↓	      determine in discussing a culture, and the site of the
␈↓"β␈↓ α←␈↓	      culture is one of them.
␈↓"β␈↓ α←␈↓	        The site of CULTURE-1 may be useful later in the
␈↓"β␈↓ α←␈↓	      consultation.  For example, it is very important in
␈↓"β␈↓ α←␈↓	      determining the identity of an organism, and is
␈↓"β␈↓ α←␈↓	      significant in determining whether or not an organism is
␈↓"β␈↓ α←␈↓	      a contaminant.

␈↓"β␈↓ α←␈↓␈↓ β?This␈α∃response␈α∃takes␈α∃account␈α∃of␈α∃the␈α∃motivation␈α∃for␈α∃automatically
␈↓ α←␈↓tracing␈α
the␈αprimary␈α
attributes: ␈αThey␈α
are␈αoften␈α
the␈αmost␈α
informative␈αpieces␈α
of
␈↓ α←␈↓information␈α↔about␈α↔the␈α↔problem.␈α⊗ In␈α↔response,␈α↔the␈α↔explanation␈α⊗routines
␈↓ α←␈↓examine␈α
the␈α
knowledge␈α
base␈α
to␈α
see␈α
how␈α
the␈α
answer␈α
may␈α
be␈α
used␈α
and␈α
select
␈↓ α←␈↓the␈αmost␈α
common␈αuses.␈α
 The␈αpiece␈α
of␈αinformation␈α
originally␈αrequested␈αis␈α
then
␈↓ α←␈↓classified␈α⊂as␈α⊂being␈α⊂``very␈α⊂important,''␈α⊂``significant,''␈α⊂or␈α⊂``relevant''␈α⊂to␈α⊃each␈α⊂of
␈↓ α←␈↓those␈α
uses,␈α
according␈α
to␈α
the␈α
number␈αof␈α
rules␈α
found.␈α
 In␈α
the␈α
example␈αabove,␈α
of
␈↓ α←␈↓all␈α∂the␈α∞rules␈α∂which␈α∞mention␈α∂culture␈α∞site␈α∂in␈α∞their␈α∂premise,␈α∞there␈α∂are␈α∂a␈α∞large
␈↓ α←␈↓number␈α∂that␈α∂conclude␈α∂about␈α⊂organism␈α∂identity␈α∂and␈α∂a␈α∂smaller␈α⊂number␈α∂that
␈↓ α←␈↓conclude about possible contaminants.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∞second␈α
special␈α∞purpose␈α∞mechanism␈α
is␈α∞employed␈α
to␈α∞make␈α∞sure␈α
that
␈↓ α←␈↓implicit␈α∪deductions␈α∩do␈α∪not␈α∩confuse␈α∪the␈α∩user.␈α∪ Consider␈α∩a␈α∪consultation␈α∩in

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[2]␈α⊂This␈α⊂is␈α⊂in␈α⊂response␈α⊂to␈α⊂the␈α⊂stylized␈α⊂presentation␈α⊂of␈α⊂information␈α⊂that␈α⊂is
␈↓ α←␈↓common␈α∃in␈α∃the␈α∃domain: ␈α∃A␈α∃clinician␈α∃is␈α∃accustomed␈α∃to␈α⊗offering␈α∃certain
␈↓ α←␈↓standard␈α∃pieces␈α∃of␈α∀information␈α∃about␈α∃each␈α∀topic.␈α∃ Tracing␈α∃the␈α∀primary
␈↓ α←␈↓attributes␈α⊃of␈α⊃each␈α⊂object␈α⊃as␈α⊃soon␈α⊂as␈α⊃the␈α⊃object␈α⊂is␈α⊃created␈α⊃means␈α⊃that␈α⊂the
␈↓ α←␈↓consultation presents a familiar format.
␈↓"β␈↓ α←␈↓␈↓48    EXPLANATION␈↓ 
#3-8␈↓

␈↓"β␈↓ α←␈↓which␈αthe␈αuser␈αindicates␈αthat␈αthe␈αculture␈αsite␈αis␈αblood.␈αIn␈αprinting␈αout␈αa␈αrule
␈↓ α←␈↓later␈α⊂on,␈α⊂the␈α⊂system␈α⊃might␈α⊂make␈α⊂reference␈α⊂to␈α⊂the␈α⊃fact␈α⊂that␈α⊂␈↓↓the␈α⊂site␈α⊃of␈α⊂the
␈↓ α←␈↓↓culture␈α
is␈αone␈α
of␈αthose␈α
sites␈αthat␈α
are␈αnormally␈α
sterile␈↓.␈α If␈α
the␈αuser␈α
asks␈αhow␈α
this
␈↓ α←␈↓fact was established, the program responds:

␈↓"β␈↓ α←␈↓	[i.e. How was it established that the site of CULTURE-1 is
␈↓"β␈↓ α←␈↓	 one of those sites that are normally sterile?]

␈↓"β␈↓ α←␈↓	You said so [question 4].
␈↓"β␈↓ α←␈↓	[Those sites that are normally sterile are: UPPER-GI, BLOOD,
␈↓"β␈↓ α←␈↓	BONE, BRAIN, CSF, LIVER, LOWER-RESP, PERICARDIUM,
␈↓"β␈↓ α←␈↓	PLEURAL-SPACE, SINUS.]


␈↓ α←␈↓Since␈α∞the␈α∞user␈α∞only␈α∞said␈α∞that␈α∞the␈α
site␈α∞was␈α∞blood,␈α∞and␈α∞may␈α∞not␈α∞be␈α∞aware␈α
of
␈↓ α←␈↓exactly␈α∂which␈α∂sites␈α∂are␈α∂sterile,␈α∂this␈α∂additional␈α∂information␈α∂is␈α∂made␈α∂clear␈α∞to
␈↓ α←␈↓him.

␈↓"β␈↓ α←␈↓␈↓α3-9    LIMITATIONS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈α⊃are␈α⊃several␈α⊃sources␈α⊃of␈α⊂limitations␈α⊃in␈α⊃the␈α⊃approach␈α⊃we␈α⊂have
␈↓ α←␈↓taken.

␈↓"β␈↓ α←␈↓␈↓α3-9-1    Restriction to a single framework␈↓
␈↓"β␈↓ α←␈↓␈↓ β?First,␈αthe␈αchoice␈αof␈αa␈αsingle␈αprimitive␈αoperation␈α(rule␈α
invocation)␈αand
␈↓ α←␈↓a␈α∂single␈α∂framework␈α∂(the␈α∂goal␈α∂tree)␈α⊂restricts␈α∂the␈α∂class␈α∂of␈α∂events␈α∂that␈α⊂can␈α∂be
␈↓ α←␈↓explained.␈α
 We␈α
have␈α
seen␈α
that␈α
behavior␈α
caused␈α
by␈α
mechanisms␈α
outside␈α
this
␈↓ α←␈↓framework␈α⊂must␈α⊂either␈α⊃be␈α⊂explained␈α⊂with␈α⊃special␈α⊂purpose␈α⊂routines␈α⊃or␈α⊂left
␈↓ α←␈↓unaccounted␈α∞for.␈α∞ In␈α∂terms␈α∞of␈α∞the␈α∂current␈α∞performance␈α∞program␈α∂(␈↓¬MYCIN␈↓),␈α∞the
␈↓ α←␈↓most␈αserious␈αexample␈αof␈αthe␈αlatter␈αis␈αdrug␈αselection.␈α While␈αsome␈αpart␈αof␈αthat
␈↓ α←␈↓process␈α
is␈α
expressed␈αin␈α
rules,␈α
much␈α
of␈αthe␈α
optimization␈α
phase␈α(minimizing␈α
the
␈↓ α←␈↓number␈α∞of␈α∞drugs␈α∂and␈α∞side␈α∞effects␈α∞and␈α∂maximizing␈α∞coverage)␈α∞is␈α∂in␈α∞complex
␈↓ α←␈↓␈↓¬LISP␈↓ code and hence inaccessible to the methods used in ␈↓¬TEIRESIAS␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?Two␈α⊃possible␈α⊃approaches␈α⊃to␈α⊃solving␈α⊃this␈α⊃are␈α⊃(␈↓↓i␈↓)␈α⊃put␈α∩everything␈α⊃in
␈↓ α←␈↓rules,␈α∂or␈α⊂(␈↓↓ii␈↓)␈α∂provide␈α∂multiple␈α⊂frameworks.␈α∂ Some␈α∂preliminary␈α⊂steps␈α∂toward
␈↓ α←␈↓the␈αformer␈αare␈αdiscussed␈αin␈αchapter␈α7,␈αwhich␈αdescribes␈αthe␈αuse␈αof␈αmeta-rules
␈↓ α←␈↓to␈α∩express␈α⊃many␈α∩types␈α⊃of␈α∩knowledge.␈α⊃ This␈α∩appears␈α⊃to␈α∩be␈α⊃only␈α∩a␈α⊃partial
␈↓ α←␈↓solution,␈α∃however.␈α∃ While␈α∃production␈α∀rules␈α∃are␈α∃a␈α∃general␈α∀computational
␈↓ α←␈↓mechanism,␈α
their␈α
utility␈α
as␈α
a␈α
representation␈α
for␈α
some␈α
tasks␈α
is␈α
questionable␈α
(see
␈↓ α←␈↓[Davis77a]␈α∞for␈α∞a␈α∞discussion␈α∞of␈α∞this␈α∞issue).␈α∞ Developing␈α∞multiple␈α
frameworks
␈↓ α←␈↓appears to be a more general solution, but has not yet been explored.

␈↓"β␈↓ α←␈↓␈↓α3-9-2    Lack of a general model of explanation␈↓
␈↓"β␈↓ α←␈↓␈↓ β?A␈αmore␈αfundamental␈αconceptual␈αlimitation␈αarises␈αout␈αof␈αthe␈αpriorities
␈↓ α←␈↓of␈α
our␈α
design␈α
criteria.␈α∞ ␈↓¬TEIRESIAS␈↓'s␈α
explanation␈α
facilities␈α
do␈α
indeed␈α∞supply␈α
an
␈↓ α←␈↓effective␈α⊗tool␈α⊗for␈α⊗discovering␈α⊗the␈α⊗basis␈α⊗for␈α⊗the␈α↔performance␈α⊗program's
␈↓ α←␈↓behavior␈α⊃and␈α⊃for␈α⊃examining␈α⊃its␈α⊃knowledge␈α⊃base.␈α⊃ They␈α⊃do␈α∩not,␈α⊃however,
␈↓ α←␈↓␈↓3-9␈↓ λsLIMITATIONS    49␈↓

␈↓"β␈↓ α←␈↓contain␈α∪a␈α∩particularly␈α∪sophisticated␈α∪model␈α∩of␈α∪explanation.␈α∪ This␈α∩becomes
␈↓ α←␈↓evident␈α∞in␈α∂several␈α∞ways.␈α∂ It␈α∞was␈α∞noted␈α∂earlier,␈α∞for␈α∂instance,␈α∞that␈α∂``why''␈α∞and
␈↓ α←␈↓``how''␈α∞are␈α∞underspecified␈α∞and␈α∞must␈α∞necessarily␈α∞be␈α∞expanded␈α∞by␈α∞␈↓¬TEIRESIAS␈↓␈α∞to
␈↓ α←␈↓avoid␈α
misinterpretation␈α
by␈α
the␈α
user␈α∞(e.g.,␈α
``Why''␈α
might␈α
mean␈α
``Why␈α∞did␈α
you
␈↓ α←␈↓ask␈α∞about␈α∞that␈α∞instead␈α∞of␈α∞...␈α∞,''␈α
or␈α∞``Why␈α∞did␈α∞you␈α∞ask␈α∞about␈α∞that␈α∞now,''␈α
etc.).
␈↓ α←␈↓These␈αother␈αinterpretations␈αare␈αvalid␈αquestions␈αabout␈αthe␈αsystem,␈αyet␈αthere␈αis
␈↓ α←␈↓no way in the current framework to answer them.
␈↓"β␈↓ α←␈↓␈↓ β?As␈α∞another␈α∞example,␈α∞there␈α∞are␈α∞certain␈α∞types␈α∞of␈α∞information␈α∞that␈α
are
␈↓ α←␈↓missing␈α∂from␈α∂the␈α⊂system␈α∂entirely,␈α∂which␈α∂are␈α⊂judged␈α∂necessary␈α∂if␈α∂we␈α⊂are␈α∂to
␈↓ α←␈↓supply␈αa␈αtruly␈αcomprehensive␈αexplanation␈αcapability.␈α Consider␈αthe␈αfollowing
␈↓ α←␈↓example.

␈↓"β␈↓ α←␈↓	Was the infection with ORGANISM-1 acquired while the patient
␈↓"β␈↓ α←␈↓	was hospitalized?
␈↓"β␈↓ α←␈↓	** ␈↓αWHY␈↓	
␈↓"β␈↓ α←␈↓	[i.e. Why is it important to determine if the infection with
␈↓"β␈↓ α←␈↓	ORGANISM-1 was acquired while the patient was hospitalized?]

␈↓"β␈↓ α←␈↓	[1.0] This will help to determine the identity of ORGANISM-1.
␈↓"β␈↓ α←␈↓	      It has already been established that
␈↓"β␈↓ α←␈↓	 [1.1] the gramstain of ORGANISM-1 is gramneg, and
␈↓"β␈↓ α←␈↓	 [1.2] the morphology of ORGANISM-1 is rod, and
␈↓"β␈↓ α←␈↓	 [1.3] the aerobicity of ORGANISM-1 is facultative,
␈↓"β␈↓ α←␈↓	      Therefore, if
␈↓"β␈↓ α←␈↓	 [1.4] the infection with ORGANISM-1 was acquired while the
␈↓"β␈↓ α←␈↓	       patient was hospitalized,
␈↓"β␈↓ α←␈↓	      Then
␈↓"β␈↓ α←␈↓	 there is weakly suggestive evidence (.2) that the identity of
␈↓"β␈↓ α←␈↓	 the organism is pseudomonas.
␈↓"β␈↓ α←␈↓	 [RULE050]

␈↓ α←␈↓A␈αnaive␈αuser␈αmight␈αnow␈αbe␈αtempted␈α
to␈αask␈αa␈αdifferent␈αsort␈αof␈α``why'': ␈α␈↓↓Why␈α
is
␈↓ α←␈↓↓it␈αtrue␈αthat␈αa␈αgram␈αnegative␈αfacultative␈αrod␈αin␈αa␈αhospital␈αsetting␈αis␈αlikely␈αto␈αbe
␈↓ α←␈↓↓a␈α∀pseudomonas?␈↓ ␈α∃The␈α∀answer␈α∀is␈α∃that␈α∀certain␈α∀sorts␈α∃of␈α∀bacteria␈α∃are␈α∀very
␈↓ α←␈↓common␈αin␈αthe␈αhospital␈αenvironment,␈α
but␈αthere␈αis␈αcurrently␈αno␈α
representation
␈↓ α←␈↓of␈αthis␈αin␈αthe␈αsystem.␈α Since␈αthe␈αsequence␈αof␈α``why''␈αquestions␈αof␈αthis␈αsort␈αcan
␈↓ α←␈↓be␈α↔continued␈α↔indefinitely␈α⊗(␈↓↓Why␈α↔are␈α↔some␈α⊗bacteria␈α↔more␈α↔common␈α↔in␈α⊗the
␈↓ α←␈↓↓hospital?␈↓,␈αetc.),␈αit␈αis␈αnot␈αunreasonable␈αto␈αcut␈αoff␈αat␈αsome␈αpoint.␈α But␈αcurrently
␈↓ α←␈↓the␈α↔question␈α↔cannot␈α_be␈α↔answered␈α↔even␈α↔once,␈α_and␈α↔in␈α↔a␈α_more␈α↔general
␈↓ α←␈↓explanation model, this should be possible.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∃lack␈α∃of␈α⊗a␈α∃user␈α∃model␈α∃in␈α⊗the␈α∃information␈α∃metric␈α⊗has␈α∃been
␈↓ α←␈↓mentioned,␈αand␈α
this␈αtoo␈α
is␈αan␈αimportant␈α
element.␈α In␈α
our␈αcurrent␈α
model,␈αtwo
␈↓ α←␈↓rules␈αwith␈αthe␈αsame␈αcertainty␈αfactor␈αhave␈αthe␈αsame␈αinformation␈αcontent,␈αyet␈αa
␈↓ α←␈↓user␈α∪who␈α∪is␈α∪familiar␈α∀with␈α∪only␈α∪the␈α∪first␈α∀of␈α∪them␈α∪will␈α∪find␈α∀much␈α∪more
␈↓ α←␈↓information in the second.
␈↓"β␈↓ α←␈↓␈↓ β?Next,␈α↔while␈α_our␈α↔approach␈α_generates␈α↔high-level␈α_explanations␈α↔by
␈↓ α←␈↓leaving␈α∂out␈α∂detail,␈α⊂there␈α∂is␈α∂another␈α∂sort␈α⊂of␈α∂abstraction␈α∂that␈α∂would␈α⊂be␈α∂very
␈↓ α←␈↓useful.␈α∃ The␈α∃system␈α∃should␈α∃be␈α∀able␈α∃to␈α∃describe␈α∃its␈α∃actions␈α∃at␈α∀different
␈↓"β␈↓ α←␈↓␈↓50    EXPLANATION␈↓ 
#3-9␈↓

␈↓"β␈↓ α←␈↓conceptual␈α∩levels,␈α∩perhaps␈α∩at␈α∩levels␈α∩ranging␈α∩from␈α∩rule␈α∩invocation␈α∪to␈α∩␈↓¬LISP␈↓
␈↓ α←␈↓interpretation,␈αto␈αfunction␈αcalls,␈αetc.␈α A␈αlimited␈αform␈αof␈αthis␈αkind␈αof␈αability␈αis
␈↓ α←␈↓described␈α∀in␈α∪[Sacerdoti77],␈α∀but␈α∪there␈α∀it␈α∪arises␈α∀from␈α∪code␈α∀that␈α∀has␈α∪been
␈↓ α←␈↓intentionally␈α∂structured␈α∞in␈α∂this␈α∞multilevel␈α∂form.␈α∞ More␈α∂interesting␈α∂would␈α∞be
␈↓ α←␈↓an␈α∀ability␈α∀to␈α∪generate␈α∀such␈α∀explanations␈α∪from␈α∀an␈α∀understanding␈α∀of␈α∪the
␈↓ α←␈↓process␈α∩at␈α⊃each␈α∩level.␈α⊃ Some␈α∩of␈α∩the␈α⊃work␈α∩on␈α⊃models␈α∩of␈α∩disease␈α⊃processes
␈↓ α←␈↓described in [Kulikowski73] may be relevant here.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈αthere␈α
is␈αthe␈α
possibility␈αof␈α
decoupling␈αexplanation␈αfrom␈α
control
␈↓ α←␈↓flow.␈α
 Section␈α
3-2␈α∞pointed␈α
out␈α
a␈α∞fundamental␈α
assumption␈α
of␈α∞our␈α
approach: 
␈↓ α←␈↓that␈α∀a␈α∪recap␈α∀of␈α∀program␈α∪actions␈α∀can␈α∀be␈α∪a␈α∀plausible␈α∀explanation.␈α∪ This
␈↓ α←␈↓assumption␈α_need␈α_not␈α_be␈α_made␈α_and␈α_explanations␈α_could␈α_be␈α_considered
␈↓ α←␈↓separately,␈α∃distinct␈α∃from␈α∃execution␈α∀(as␈α∃in␈α∃[Brown75]).␈α∃ This␈α∃is␈α∀common
␈↓ α←␈↓human␈α∂behavior--the␈α∂account␈α∂someone␈α∂gives␈α∂of␈α∂how␈α∂he␈α∂solved␈α∂a␈α∂complex
␈↓ α←␈↓problem␈α∂may␈α∂be␈α∞quite␈α∂different␈α∂from␈α∞a␈α∂simple␈α∂review␈α∞of␈α∂his␈α∂actions.␈α∞ The
␈↓ α←␈↓difference␈αis␈αoften␈αmore␈αthan␈αjust␈αleaving␈αout␈αthe␈αdead␈αends␈αor␈αomitting␈αthe
␈↓ α←␈↓details␈αof␈αthe␈αsolution.␈α Solving␈αthe␈αproblem␈αoften␈αproduces␈αnew␈αinsights␈αand
␈↓ α←␈↓shows␈α
results␈α
in␈α
a␈α
totally␈α
different␈α
view,␈α
one␈α
which␈α
often␈α
admits␈α
a␈αmuch␈α
more
␈↓ α←␈↓compact␈α∞explanation.␈α
 There␈α∞is␈α∞greater␈α
flexibility␈α∞from␈α
an␈α∞approach␈α∞of␈α
this
␈↓ α←␈↓kind␈α∞since␈α∞it␈α∞can␈α∞be␈α∞used␈α∞to␈α∞supply␈α∞a␈α∞wider␈α∞variety␈α∞of␈α∞explanations;␈α∂but␈α∞it
␈↓ α←␈↓also␈α∩requires␈α∩a␈α∩new␈α∩and␈α∩separate␈α∩basis␈α∩for␈α∩generating␈α∪explanations␈α∩and,
␈↓ α←␈↓hence, is more difficult.
␈↓"β␈↓ α←␈↓␈↓ β?All␈α
of␈α
these␈α
examples␈α
point␈α
the␈α
way␈α
to␈α
useful␈α
extensions␈α
to␈α␈↓¬TEIRESIAS␈↓
␈↓ α←␈↓that would give it a far more general and powerful explanation capability.

␈↓"β␈↓ α←␈↓␈↓α3-9-3    Lack of ability to represent control structures␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
third,␈α∞and␈α
most␈α∞relevant,␈α
limitation␈α∞is␈α
the␈α∞lack␈α
of␈α∞a␈α
substantive
␈↓ α←␈↓ability␈α∪to␈α∪represent␈α∪control␈α∀structures.␈α∪ In␈α∪the␈α∪current␈α∪system,␈α∀the␈α∪closest
␈↓ α←␈↓approximation␈αcomes␈αfrom␈αthe␈αuse␈αof␈αthe␈αgoal-tree␈αframework␈αas␈αa␈αmodel␈αof
␈↓ α←␈↓the␈α
control␈αstructure.␈α
 But␈α
this␈α``model''␈α
exists␈α
only␈αimplicitly,␈α
expressed␈αby␈α
the
␈↓ α←␈↓organization␈α⊂of␈α⊂code␈α⊂in␈α⊂the␈α⊂explanation␈α⊂facility␈α⊂rather␈α⊂than␈α⊂as␈α⊂a␈α∂separate
␈↓ α←␈↓entity.␈α Shortcomings␈αare␈αevidenced␈αby␈α
the␈αfact␈αthat␈αalmost␈αany␈α
alteration␈αof
␈↓ α←␈↓the␈α∪control␈α∩structure␈α∪of␈α∩the␈α∪performance␈α∩program␈α∪requires␈α∪an␈α∩analogous
␈↓ α←␈↓recoding␈α∀of␈α∪␈↓¬TEIRESIAS␈↓'s␈α∀explanation␈α∪routines.␈α∀ If␈α∪there␈α∀were␈α∀instead␈α∪some
␈↓ α←␈↓representation␈α⊂of␈α⊂the␈α⊂control␈α⊂structure␈α⊂that␈α⊂the␈α⊂explanation␈α⊃routines␈α⊂could
␈↓ α←␈↓examine,␈α∩the␈α∩system␈α∩would␈α∩be␈α∩much␈α∩more␈α∩flexible.␈α∩ The␈α∩general␈α⊃scheme
␈↓ α←␈↓would be as pictured below:
␈↓ α←␈↓␈↓3-9␈↓ λsLIMITATIONS    51␈↓



␈↓"␈↓ α←␈↓∧⊂αααααααααααααα⊃     ⊂ααααααααααα⊃     ⊂ααααααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧~representation~     ~ execution ~     ~  knowledge of what  ~
␈↓"␈↓ α←␈↓∧~of control    ~     ~   trace   ~     ~     constitutes     ~
␈↓"␈↓ α←␈↓∧~structure     ~     %ααααααααααα$     ~   an explanation    ~
␈↓"␈↓ α←␈↓∧%αααααααααααααα$α ⊃        ~        ⊂ α%ααααααααααααααααααααα$

␈↓"␈↓ α←␈↓∧                  ~        ~        ~

␈↓"␈↓ α←␈↓∧                  ↓        ↓        ↓
␈↓"␈↓ α←␈↓∧                  ⊂ααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧                  ~   explanation   ~
␈↓"␈↓ α←␈↓∧                  ~    generator    ~
␈↓"␈↓ α←␈↓∧                  %ααααααααααααααααα$
␈↓"␈↓ α←␈↓∧                           |
␈↓"␈↓ α←␈↓∧                           ↓
␈↓"␈↓ α←␈↓∧            explanation of program behavior

␈↓"␈↓ α←␈↓α␈↓ β∩Fig. 3-5.    A more sophisticated explanation-generation scheme.    


␈↓ α←␈↓Given␈αsome␈αrepresentation␈αof␈αthe␈αcontrol␈αstructure␈αand␈αa␈αbody␈αof␈αknowledge
␈↓ α←␈↓about␈α⊃what␈α⊃an␈α⊃explanation␈α∩is,␈α⊃␈↓¬TEIRESIAS␈↓␈α⊃would␈α⊃generate␈α∩a␈α⊃comprehensible
␈↓ α←␈↓account␈α
of␈α
performance␈α
program␈αbehavior.␈α
 Supplying␈α
either␈α
of␈α
these␈αseems
␈↓ α←␈↓to␈αbe␈αa␈αformidable␈α
but␈αinteresting␈αprospect.␈α The␈α
fact␈αthat␈αa␈αprogrammer␈α
can
␈↓ α←␈↓scan␈α∂a␈α∞strange␈α∂piece␈α∂of␈α∞code␈α∂and␈α∞then␈α∂explain␈α∂the␈α∞behavior␈α∂of␈α∂a␈α∞program
␈↓ α←␈↓executing␈α⊃it,␈α∩suggests␈α⊃that␈α∩the␈α⊃task␈α∩is␈α⊃at␈α∩least␈α⊃plausible.␈α∩ Formalizing␈α⊃the
␈↓ α←␈↓knowledge␈α⊂of␈α⊂what␈α⊂it␈α⊂means␈α⊂to␈α⊂``explain''␈α⊂does␈α⊂not␈α⊂seem␈α⊂to␈α⊃have␈α⊂received
␈↓ α←␈↓extensive␈α_attention.␈α_ Some␈α_of␈α_the␈α_work␈α_on␈α_affects␈α_and␈α_intensions␈α↔by
␈↓ α←␈↓[Faught74]␈α∞may␈α
be␈α∞relevant,␈α
as␈α∞may␈α
be␈α∞the␈α
conceptual␈α∞dependency␈α∞work␈α
in
␈↓ α←␈↓[Reiger74] and the work on computer-aided instruction in [Brown75].
␈↓"β␈↓ α←␈↓␈↓ β?Some␈α∪small␈α∪steps␈α∪toward␈α∪representing␈α∪the␈α∪control␈α∪structure␈α∪in␈α∪an
␈↓ α←␈↓accessible␈α⊂form␈α⊃are␈α⊂described␈α⊂in␈α⊃chapter␈α⊂7.␈α⊂ There␈α⊃we␈α⊂explore␈α⊂the␈α⊃use␈α⊂of
␈↓ α←␈↓meta-rules␈α∂to␈α∂express␈α∂certain␈α⊂aspects␈α∂of␈α∂the␈α∂performance␈α⊂program's␈α∂control
␈↓ α←␈↓structure,␈α∃and␈α∃hence␈α∃render␈α⊗them␈α∃accessible␈α∃to␈α∃the␈α⊗current␈α∃explanation
␈↓ α←␈↓mechanism.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α
more␈αsophisticated␈α
form␈αof␈α
representation␈αis␈α
needed,␈αhowever.␈α
 The
␈↓ α←␈↓source␈α
code␈α
of␈α
a␈α∞program␈α
is␈α
not␈α
a␈α∞particularly␈α
good␈α
choice␈α
for␈α∞two␈α
reasons.
␈↓ α←␈↓First,␈α∞it␈α∂carries␈α∞too␈α∞much␈α∂implementation␈α∞detail.␈α∞ It␈α∂may␈α∞be␈α∞useful␈α∂to␈α∞know
␈↓ α←␈↓only␈αthat␈αa␈αpiece␈α
of␈αcode␈αis,␈αfor␈αinstance,␈α
an␈αiterative␈αloop,␈αand␈αthe␈α
detail␈αof
␈↓ α←␈↓how␈α∞it␈α∞is␈α∂implemented␈α∞often␈α∞only␈α∞confuses␈α∂the␈α∞issue.␈α∞ Second,␈α∞it␈α∂carries␈α∞too
␈↓ α←␈↓little␈α
intentional␈α
information.␈α
 ␈↓¬LISP␈↓'s␈α∞␈↓	CAR␈↓␈α
operation,␈α
for␈α
example,␈α∞can␈α
usually
␈↓ α←␈↓be␈α∪read␈α∪as␈α∩␈↓↓first␈α∪element␈↓␈α∪of␈α∩a␈α∪list;␈α∪but␈α∩inside␈α∪a␈α∪loop,␈α∩it␈α∪might␈α∪be␈α∩more
␈↓ α←␈↓appropriately viewed as ␈↓↓next element␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αtask␈α
requires␈αsome␈α
way␈αof␈α
suppressing␈αimplementation␈αdetail␈α
and
␈↓ α←␈↓emphasizing␈α∀intentions.␈α∀ There␈α∪is␈α∀a␈α∀common␈α∪style␈α∀of␈α∀programming␈α∪that
␈↓ α←␈↓␈↓52    EXPLANATION␈↓ 
#3-9␈↓

␈↓"β␈↓ α←␈↓accomplishes␈αsome␈αof␈αthis.␈α
 It␈αemphasizes␈αthe␈αuse␈α
of␈αextended␈αdata␈αtypes␈α
(e.g.,
␈↓ α←␈↓record␈αstructures)␈α
and␈αmacros␈α
(or␈αfunctions)␈α
designed␈αfor␈α
the␈αtask␈α
at␈αhand.␈α
 It
␈↓ α←␈↓may␈αinclude␈αtwo␈αfunctions␈αcalled␈α␈↓	FIRST␈↓␈αand␈α␈↓	NEXT␈↓,␈αfor␈αinstance,␈αeven␈αthough
␈↓ α←␈↓both␈α∞perform␈α∞the␈α
same␈α∞␈↓	CAR␈↓␈α∞operation.␈α∞ Something␈α
along␈α∞this␈α∞line␈α∞has␈α
been
␈↓ α←␈↓suggested␈α∂in␈α⊂[Hewitt71],␈α∂but␈α⊂the␈α∂emphasis␈α∂there␈α⊂is␈α∂on␈α⊂program␈α∂correctness
␈↓ α←␈↓and␈α→automatic␈α_programming.␈α→ The␈α_point␈α→here␈α_is␈α→simply␈α_to␈α→give␈α_the
␈↓ α←␈↓programmer␈α
a␈α
way␈α
of␈α
expressing␈α
his␈α
(perhaps␈α
informal)␈α
intentions␈α∞for␈α
each
␈↓ α←␈↓section␈α
of␈αcode␈α
so␈αthat␈α
the␈αprogram␈α
can␈αlater␈α
be␈αexplained.␈α
 Note␈α
that␈αthese
␈↓ α←␈↓intentions␈αmight␈αinclude␈α
not␈αonly␈αdescriptions␈α
of␈αcontrol␈αstructure␈α
but␈αmight
␈↓ α←␈↓go␈α∂on␈α∂to␈α∂specify␈α∂many␈α∞other␈α∂things␈α∂about␈α∂the␈α∂code: ␈α∂design␈α∞considerations,
␈↓ α←␈↓algorithm␈αchoices,␈αand␈αperhaps␈αother␈αkinds␈αof␈αinformation␈αof␈αthe␈αsort␈αfound
␈↓ α←␈↓in␈α≠well-written␈α~documentation.␈α≠ Thus,␈α≠where␈α~the␈α≠efforts␈α≠in␈α~program
␈↓ α←␈↓correctness␈α∃have␈α∀concentrated␈α∃on␈α∀describing␈α∃precisely␈α∀``what''␈α∃a␈α∀program
␈↓ α←␈↓should␈αdo,␈α
intentional␈αinformation␈α
might␈αinclude␈α
descriptions␈αof␈α
the␈α``how''␈α
of
␈↓ α←␈↓program␈α
design.␈α
 The␈α∞work␈α
in␈α
[Goldstein74]␈α∞suggests␈α
the␈α
multiple␈α∞uses␈α
that
␈↓ α←␈↓might␈α⊂be␈α⊂made␈α⊂of␈α⊂such␈α⊂annotations: ␈α⊂That␈α⊂system␈α⊂based␈α⊂its␈α⊂debugging␈α∂of
␈↓ α←␈↓simple␈α
programs␈α
on␈α
an␈α
understanding␈αof␈α
common␈α
error␈α
types␈α
plus␈αremarks␈α
in
␈↓ α←␈↓the code that indicated the programmer's ``plan'' for solving the problem.
␈↓"β␈↓ α←␈↓␈↓ β?Flowcharts␈αare␈αa␈α
second␈αpossibility;␈αsome␈α
text-oriented␈αrepresentation
␈↓ α←␈↓of␈αthem␈αmight␈αprove␈αuseful.␈α In␈αan␈αanalogy␈αto␈αsome␈αof␈αthe␈αwork␈αon␈αprogram
␈↓ α←␈↓correctness,␈αthey␈αmight␈αbe␈αannotated␈αwith␈α``motivation␈αconditions,''␈αproviding,
␈↓ α←␈↓thereby,␈α∞both␈α∞a␈α∞representation␈α∂of␈α∞the␈α∞control␈α∞structure␈α∞and␈α∂the␈α∞information
␈↓ α←␈↓necessary␈α∞to␈α∂explain␈α∞it.␈α∞ Both␈α∂of␈α∞these␈α∞are␈α∂clearly␈α∞very␈α∞speculative;␈α∂there␈α∞is
␈↓ α←␈↓much room for additional work.

␈↓"β␈↓ α←␈↓␈↓α3-9-4    Other communication media␈↓
␈↓"β␈↓ α←␈↓␈↓ β?We␈α⊂have␈α⊂not␈α⊃yet␈α⊂explored␈α⊂any␈α⊂graphics-oriented␈α⊃explanations,␈α⊂but
␈↓ α←␈↓Fig.␈α⊂3-1␈α∂illustrates␈α⊂that␈α∂they␈α⊂might␈α⊂be␈α∂very␈α⊂useful.␈α∂ The␈α⊂tree␈α⊂shown␈α∂there
␈↓ α←␈↓makes␈α
clear␈α
what␈αotherwise␈α
takes␈α
many␈α
lines␈αof␈α
text.␈α
 Considering␈αthe␈α
natural
␈↓ α←␈↓interpretation␈α
of␈α
the␈α
performance␈αprogram's␈α
control␈α
structure␈α
in␈αgraphic␈α
form
␈↓ α←␈↓and␈α∩the␈α∩desire␈α∩to␈α∪be␈α∩as␈α∩brief␈α∩as␈α∩possible,␈α∪it␈α∩appears␈α∩to␈α∩be␈α∪an␈α∩excellent
␈↓ α←␈↓medium.

␈↓"β␈↓ α←␈↓␈↓α3-10    TRACE OF SYSTEM PERFORMANCE:  EXPLANATIONS FOR
␈↓ α←␈↓α␈↓ β3SYSTEM DEBUGGING␈↓
␈↓"β␈↓ α←␈↓␈↓ ε<And you are wrong if you believe that one. 
␈↓"β␈↓ α←␈↓␈↓ 	gline 554
␈↓"β␈↓ α←␈↓␈↓ β?Work␈α_on␈α_giving␈α_a␈α_program␈α_the␈α_ability␈α_to␈α_explain␈α→its␈α_actions
␈↓ α←␈↓([Shortliffe75a])␈α1was␈α1originally␈α0motivated␈α1by␈α1concerns␈α0about
␈↓ α←␈↓comprehensibility:␈α
Consultation␈α
programs␈α
are␈α
unlikely␈α
to␈α
be␈α
accepted␈αby␈α
users
␈↓ α←␈↓if␈αthey␈α
function␈αas␈α
``black␈αboxes''␈αthat␈α
simply␈αprint␈α
their␈αfinal␈αanswers.␈α
 Much
␈↓ α←␈↓of the work described earlier in this chapter is in the same vein.
␈↓"β␈↓ α←␈↓␈↓ β?But␈α↔there␈α↔is␈α↔another,␈α⊗different,␈α↔and␈α↔very␈α↔useful␈α↔application␈α⊗of
␈↓ α←␈↓explanation:␈α∪program␈α∩debugging,␈α∪in␈α∩particular␈α∪as␈α∩a␈α∪means␈α∩of␈α∪setting␈α∩an
␈↓ α←␈↓appropriate␈α⊂context␈α⊃for␈α⊂knowledge␈α⊂acquisition.␈α⊃ As␈α⊂an␈α⊂introduction␈α⊃to␈α⊂the
␈↓ α←␈↓␈↓3-10␈↓ ε!EXPLANATIONS FOR SYSTEM DEBUGGING    53␈↓

␈↓"β␈↓ α←␈↓knowledge␈α∂acquisition␈α∂sections␈α⊂that␈α∂follow,␈α∂an␈α∂example␈α⊂below␈α∂demonstrates
␈↓ α←␈↓how␈α∞the␈α∂explanation␈α∞routines␈α∂in␈α∞␈↓¬TEIRESIAS␈↓␈α∂have␈α∞been␈α∂designed␈α∞to␈α∂allow␈α∞the
␈↓ α←␈↓domain expert to track down the source of an error in the knowledge base.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α
of␈αthe␈α
options␈αavailable␈α
to␈αthe␈α
expert␈αis␈α
having␈αthe␈α
system␈αstop
␈↓ α←␈↓after␈α
it␈α
has␈α
presented␈α∞its␈α
diagnosis,␈α
to␈α
give␈α
him␈α∞a␈α
chance␈α
to␈α
comment␈α∞on␈α
it.
␈↓ α←␈↓This␈α
pause␈α
provides␈α
a␈α
natural␈α
starting␈α
place␈α
for␈α
evaluation␈α
and␈α
debugging.
␈↓ α←␈↓Once␈α∞the␈α∂bug␈α∞is␈α∞found,␈α∂the␈α∞acquisition␈α∞session,␈α∂described␈α∞in␈α∂later␈α∞chapters,
␈↓ α←␈↓then goes on to repair the problem by teaching the system the new rule.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α~example␈α~presented␈α→here␈α~will␈α~be␈α→carried␈α~on␈α~through␈α→the
␈↓ α←␈↓knowledge␈αacquisition␈αchapters␈αand␈αdeveloped␈αin␈αdetail.␈α In␈αorder␈α
to␈αpresent
␈↓ α←␈↓a␈α
single␈α
example␈α∞that␈α
would␈α
include␈α∞all␈α
the␈α
features␈α∞to␈α
be␈α
displayed,␈α∞it␈α
was
␈↓ α←␈↓necessary␈α
to␈α
create␈α
a␈α
bug␈α
by␈α
removing␈α
a␈α
rule␈α
from␈α
the␈α
performance␈α
program's
␈↓ α←␈↓knowledge␈α
base␈α
and␈α
eliminating␈α
all␈αreferences␈α
to␈α
it.␈α
 The␈α
example␈α
is␈αspecial
␈↓ α←␈↓in␈αno␈αother␈αsense,␈αhowever,␈αand␈αwith␈αthe␈αobvious␈αexceptions␈αof␈αfont␈αchanges,
␈↓ α←␈↓annotations,␈α∞and␈α∞slight␈α∞reformatting,␈α∂the␈α∞interaction␈α∞appears␈α∞just␈α∞as␈α∂it␈α∞does
␈↓ α←␈↓here.  For completeness, the entire consultation is presented below.

␈↓ α←␈↓	***************************************************************

␈↓ α←␈↓	MYCIN  24-NOV-76 ...

␈↓ α←␈↓	Special options (type ? for help):
␈↓ α←␈↓	** ␈↓αRD SAVE␈↓	
␈↓ α←␈↓	Hello, Dr. Davis

␈↓ α←␈↓↓␈↓ βW{The␈α⊂expert␈α⊂identifies␈α⊂himself␈α⊂to␈α⊂the␈α⊂system␈α⊂and␈α⊂indicates␈α⊂that␈α∂he
␈↓ α←␈↓↓␈↓ βWwants a record of this consultation kept in the case library.}

␈↓ α←␈↓	This will be patient 281.
␈↓ α←␈↓	If you would care to enter a brief clinical summary of your
␈↓ α←␈↓	patient's case, to be used for easy future reference to this
␈↓ α←␈↓	file, please do so now (type as many lines as required,
␈↓ α←␈↓	terminate with a blank line):

␈↓ α←␈↓	**␈↓αCLASSIC CASE OF GRAM NEGATIVE ROD INFECTION FOR A ␈↓	
␈↓ α←␈↓	**␈↓αPATIENT WITH A NON-NOSOCOMIAL DISEASE␈↓	
␈↓ α←␈↓	**

␈↓ α←␈↓	--------PATIENT-281--------
␈↓ α←␈↓	1) Patient's name: (first-last)
␈↓ α←␈↓	** ␈↓αALBERT HALL␈↓	
␈↓ α←␈↓	2) Sex:
␈↓ α←␈↓	** ␈↓αMALE␈↓	
␈↓ α←␈↓	3) Age:
␈↓ α←␈↓	** ␈↓α34␈↓	
␈↓ α←␈↓	4) Have you been able to obtain positive microbiological
␈↓ α←␈↓	   information about a possiBle infection of Albert Hall?
␈↓ α←␈↓	** ␈↓αYES␈↓	
␈↓"β␈↓ α←␈↓␈↓54    EXPLANATION␈↓ 
∪3-10␈↓

␈↓"β␈↓ α←␈↓	 --------INFECTION-1--------
␈↓ α←␈↓	 5) What is the infection?
␈↓ α←␈↓	 ** ␈↓αPRIMARY-BACTEREMIA␈↓	
␈↓ α←␈↓	 6) Please give the date and time when signs or symptoms of
␈↓ α←␈↓	    the primary-bacteremia (INFECTION-1) first appeared.
␈↓ α←␈↓	 ** ␈↓α15 DECEMBER 77 12:00␈↓	
␈↓ α←␈↓	 The most recent culture yielding positive microbiological
␈↓ α←␈↓	 information  about the primary-bacteremia (INFECTION-1) will
␈↓ α←␈↓	 be referred to as:

␈↓ α←␈↓	  --------CULTURE-1--------
␈↓ α←␈↓	  7) From what site was the specimen for CULTURE-1 taken?
␈↓ α←␈↓	  ** ␈↓αBLOOD␈↓	
␈↓ α←␈↓	  8) Please give the date and time when this blood culture
␈↓ α←␈↓	     (CULTURE-1) was obtained. (mo/da/yr time)
␈↓ α←␈↓	  ** ␈↓α16 DECEMBER 77 12:00␈↓	
␈↓ α←␈↓	  The first significant organism from this blood culture
␈↓ α←␈↓	  (CULTURE-1) will be referred to as:

␈↓ α←␈↓	   --------ORGANISM-1--------
␈↓ α←␈↓	   9) Enter the identity of ORGANISM-1:
␈↓ α←␈↓	   ** ␈↓αUNKNOWN␈↓	
␈↓ α←␈↓	   10) Is ORGANISM-1 a rod or coccus (etc.):
␈↓ α←␈↓	   ** ␈↓αROD␈↓	
␈↓ α←␈↓	   11) The gram stain of ORGANISM-1:
␈↓ α←␈↓	   ** ␈↓αGRAMNEG␈↓	
␈↓ α←␈↓	   12) Did ORGANISM-1 grow in the aerobic bottle?
␈↓ α←␈↓	   ** ␈↓αYES␈↓	
␈↓ α←␈↓	   13) Did ORGANISM-1 grow in the anaerobic bottle?
␈↓ α←␈↓	   ** ␈↓αYES␈↓	
␈↓ α←␈↓	   14) Is the patient's illness with ORGANISM-1 a hospital-
␈↓ α←␈↓	       acquired infection?
␈↓ α←␈↓	   ** ␈↓αNO␈↓	
␈↓ α←␈↓	   15) Have there been POSITIVE cultures yielding organisms
␈↓ α←␈↓	       about which you will NOT be seeking advice?
␈↓ α←␈↓	   ** ␈↓αNO␈↓	
␈↓ α←␈↓	   16) Is Albert Hall a burn patient?
␈↓ α←␈↓	   ** ␈↓αNO␈↓	
␈↓ α←␈↓	   17) Do you consider Albert Hall to be a compromised host
␈↓ α←␈↓	       (alcoholic, immunosuppressed, etc.)?
␈↓ α←␈↓	   ** ␈↓αNO␈↓	
␈↓ α←␈↓	   18) What is the suspected portal of entry of ORGANISM-1
␈↓ α←␈↓	       into this sterile site?
␈↓ α←␈↓	   ** ␈↓αGI␈↓	
␈↓ α←␈↓	   19) Has Albert Hall had any G-U surgery or manipulation of
␈↓ α←␈↓	       the urinary tract (eg. catheter, cystoscopy, etc.)?
␈↓ α←␈↓	   ** ␈↓αNO␈↓	
␈↓ α←␈↓	   20) Including this blood culture (CULTURE-1), how many times
␈↓ α←␈↓	       during this series of cultures was this site cultured?
␈↓ α←␈↓	   ** ␈↓α3␈↓	
␈↓ α←␈↓	   21) In this series, how many of the cultures from this
␈↓ α←␈↓	       site were positive for ORGANISM-1?
␈↓ α←␈↓	   ** ␈↓α3␈↓	
␈↓"β␈↓ α←␈↓␈↓3-10␈↓ ε!EXPLANATIONS FOR SYSTEM DEBUGGING    55␈↓

␈↓"β␈↓ α←␈↓	  22) Any other significant organisms isolated from this blood
␈↓ α←␈↓	      culture (CULTURE-1)?
␈↓ α←␈↓	  ** ␈↓αNO␈↓	
␈↓ α←␈↓	 23) Did any additional cultures yield positive microbiological
␈↓ α←␈↓	     information which you believe is associated with the
␈↓ α←␈↓	     primary-bacteremia (INFECTION-1)?
␈↓ α←␈↓	 ** ␈↓αNO␈↓	
␈↓ α←␈↓	24) Does Albert Hall have any other infection about which you
␈↓ α←␈↓	    have obtained positive microbiological information?
␈↓ α←␈↓	** ␈↓αNO␈↓	
␈↓ α←␈↓	25) Do you suspect an infection at a site which you have been
␈↓ α←␈↓	    unable to culture, or from which all cultures were
␈↓ α←␈↓	    negative?
␈↓ α←␈↓	** ␈↓αNO␈↓	
␈↓ α←␈↓	26) Is Albert Hall currently receiving therapy with any
␈↓ α←␈↓	    antimicrobial agent?
␈↓ α←␈↓	** ␈↓αNO␈↓	
␈↓ α←␈↓	27) During the present illness, has Albert Hall been treated
␈↓ α←␈↓	    with an antimicrobial agent which is no longer being
␈↓ α←␈↓	    administered?
␈↓ α←␈↓	** ␈↓αNO␈↓	

␈↓ α←␈↓	My therapy recommendation will be based on the following
␈↓ α←␈↓	probable infection(s) and potential causative organism(s):

␈↓ α←␈↓	INFECTION-1 is PRIMARY-BACTEREMIA

␈↓ α←␈↓	    <Item 1>  E.COLI [ORGANISM-1]
␈↓ α←␈↓	    <Item 2>  PSEUDOMONAS-AERUGINOSA [ORGANISM-1]


␈↓ α←␈↓	Are all these identities correct?
␈↓ α←␈↓	++** ␈↓αN␈↓	

␈↓ α←␈↓↓␈↓ βW{At␈αthis␈αpoint␈α␈↓¬TEIRESIAS␈↓↓␈αsteps␈αin,␈αasks␈αfor␈αthe␈αexpert's␈αopinion,␈αand␈αis
␈↓ α←␈↓↓␈↓ βWtold␈α∀that␈α∀something␈α∀is␈α∃wrong.␈α∀ From␈α∀here␈α∀on,␈α∀the␈α∃interaction␈α∀is
␈↓ α←␈↓↓␈↓ βWbetween␈α∂the␈α∂expert␈α∂and␈α∞␈↓¬TEIRESIAS␈↓↓,␈α∂with␈α∂the␈α∂performance␈α∂program␈α∞in
␈↓ α←␈↓↓␈↓ βWthe background being manipulated by ␈↓¬TEIRESIAS␈↓↓}.

␈↓ α←␈↓	Are there any identities missing?
␈↓ α←␈↓	++** ␈↓αY␈↓	
␈↓ α←␈↓	Which one(s)?
␈↓ α←␈↓	++** ␈↓αKLEBSIELLA-PNEUMONIAE PROTEUS-NON-MIRABILIS␈↓	
␈↓ α←␈↓	Are there any identities that shouldn't have appeared?
␈↓ α←␈↓	++** ␈↓αY␈↓	
␈↓ α←␈↓	Which one(s)?
␈↓ α←␈↓	++** ␈↓αPSEUDOMONAS-AERUGINOSA␈↓	

␈↓ α←␈↓	We'll start with the PSEUDOMONAS-AERUGINOSA first.
␈↓ α←␈↓␈↓56    EXPLANATION␈↓ 
∪3-10␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βW{Heuristics␈α∞derived␈α∞from␈α∞our␈α
experience␈α∞in␈α∞debugging␈α∞are␈α∞built␈α
into
␈↓ α←␈↓↓␈↓ βWthe␈αsystem.␈α They␈αsuggest␈αthat␈αthe␈αbest␈αapproach␈αis␈α(a)␈αto␈αtrack␈αdown
␈↓ α←␈↓↓␈↓ βWone␈αproblem␈αat␈αa␈αtime,␈αand␈α(b)␈αto␈αstart␈αwith␈αan␈αanswer␈αthat␈αshouldn't
␈↓ α←␈↓↓␈↓ βWhave␈αappeared,␈αif␈αthere␈αis␈αone.␈α Experience␈αhas␈αsuggested␈αthat␈αvalues
␈↓ α←␈↓↓␈↓ βWwhich␈α
appear␈α∞inappropriately␈α
are␈α∞often␈α
the␈α
result␈α∞of␈α
a␈α∞bug␈α
causing
␈↓ α←␈↓↓␈↓ βWseveral␈α
problems;␈α
so␈αcuring␈α
one␈α
often␈α
removes␈αa␈α
number␈α
of␈αproblems␈α
at
␈↓ α←␈↓↓␈↓ βWonce.}

␈↓ α←␈↓	Do you want to
␈↓ α←␈↓	      1] track down the problem yourself, or
␈↓ α←␈↓	      2] allow me to guide the debugging, or
␈↓ α←␈↓	      3] ignore it for now and fix it another time?
␈↓ α←␈↓	Type 1, 2 or 3
␈↓ α←␈↓	++** ␈↓α1␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈α→first␈α→option␈α~allows␈α→the␈α→expert␈α~to␈α→use␈α→all␈α~the␈α→standard
␈↓ α←␈↓↓␈↓ βWexplanation␈α∞facilities␈α∞and␈α∞will␈α∂be␈α∞demonstrated␈α∞here.␈α∞ The␈α∂second␈α∞is
␈↓ α←␈↓↓␈↓ βWbased␈α_on␈α↔an␈α_analysis␈α_of␈α↔the␈α_control␈α↔structure␈α_that␈α_drives␈α↔the
␈↓ α←␈↓↓␈↓ βWconsultation␈αand␈αis␈αexplored␈αin␈αthe␈αnext␈αchapter.␈α The␈αfinal␈αoption␈αis
␈↓ α←␈↓↓␈↓ βWfor␈α
human␈α
engineering␈α
purposes␈α
and␈α
allows␈α
the␈α
expert␈α
to␈α
sidestep␈α
the
␈↓ α←␈↓↓␈↓ βWwhole process if he should decide to do so.}

␈↓ α←␈↓	The following rules were used in deducing that the identity of
␈↓ α←␈↓	          ORGANISM-1 is pseudomonas-aeruginosa

␈↓ α←␈↓	RULE184
␈↓ α←␈↓	-------
␈↓ α←␈↓	Since [1.1] the category of ORGANISM-1 is not known
␈↓ α←␈↓	      [1.2] the gram stain of ORGANISM-1 is gramneg
␈↓ α←␈↓	      [1.3] the morphology of ORGANISM-1 is rod
␈↓ α←␈↓	      [1.4] the aerobicity of ORGANISM-1 is facultative

␈↓ α←␈↓	 There is weakly suggestive evidence (.3) that the identity of
␈↓ α←␈↓	         ORGANISM-1 is pseudomonas-aeruginosa

␈↓ α←␈↓	Now use the standard explanation commands (WHY HOW, etc.) to
␈↓ α←␈↓	find the bug.  When you've discovered what you think the
␈↓ α←␈↓	problem is, type "OK", and we can try to fix it

␈↓ α←␈↓	      ** ␈↓αHOW [1.1]␈↓	

␈↓ α←␈↓	[i.e. HOW was it established that the category of ORGANISM-1
␈↓ α←␈↓	 is not known?]

␈↓ α←␈↓	There were no rules successfully invoked to deduce the category
␈↓ α←␈↓	of ORGANISM-1.  The following rules failed:
␈↓ α←␈↓	      [2.1] RULE050
␈↓ α←␈↓	      [2.2] RULE116
␈↓ α←␈↓␈↓3-10␈↓ ε!EXPLANATIONS FOR SYSTEM DEBUGGING    57␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βW{When␈α⊂the␈α∂expert␈α⊂is␈α∂tracking␈α⊂down␈α⊂a␈α∂bug,␈α⊂the␈α∂answer␈α⊂to␈α⊂a␈α∂``how''
␈↓ α←␈↓↓␈↓ βWquestion␈α
includes␈α
an␈α
indication␈α
of␈α
rules␈α
that␈α
might␈α
have␈αbeen␈α
executed
␈↓ α←␈↓↓␈↓ βWbut␈αwhich␈αfailed.␈α This␈αis␈αnot␈αnormally␈αuseful␈αto␈αthe␈αaverage␈αobserver
␈↓ α←␈↓↓␈↓ βWand so is omitted in other circumstances.}

␈↓ α←␈↓	     ** ␈↓αHOW [2.1]␈↓	

␈↓ α←␈↓	It was established that
␈↓ α←␈↓	      [3.1] the gram stain of ORGANISM-1 is gramneg, and
␈↓ α←␈↓	      [3.2] the morphology of ORGANISM-1 is rod, and
␈↓ α←␈↓	      [3.3] the aerobicity of ORGANISM-1 is facultative
␈↓ α←␈↓	But it was never established that
␈↓ α←␈↓	      [3.4] the infection with ORGANISM-1 was acquired while
␈↓ α←␈↓	            the patient was hospitalized
␈↓ α←␈↓	[And the conclusion could not be made that
␈↓ α←␈↓	 there is strongly suggestive evidence (.7) that the category
␈↓ α←␈↓	 of ORGANISM-1 is enterobacteriaceae]
␈↓ α←␈↓	[RULE050]


␈↓ α←␈↓	    ** ␈↓αHOW [2.2]␈↓	

␈↓ α←␈↓	It was established that
␈↓ α←␈↓	      [4.1] the identity of the organism is not known,
␈↓ α←␈↓	But it was never established that
␈↓ α←␈↓	      [4.2] the gram stain of the organism is not known,
␈↓ α←␈↓	      [4.3] the morphology of the organism is not known, and
␈↓ α←␈↓	      [4.4] the site of the culture is csf, and
␈↓ α←␈↓	      [4.5] the infection is meningitis, and
␈↓ α←␈↓	      [4.6] the age (in years) of the patient is less than or
␈↓ α←␈↓	            equal to .17
␈↓ α←␈↓	[And the conclusion could not be made that
␈↓ α←␈↓	 there is weakly suggestive evidence (.3) that the category
␈↓ α←␈↓	 of ORGANISM-1 is enterobacteriaceae]
␈↓ α←␈↓	[RULE116]

␈↓ α←␈↓	    ** ␈↓αOK␈↓	

␈↓ α←␈↓	Good, now do you want to:
␈↓ α←␈↓	   1] give me a new rule
␈↓ α←␈↓	   2] modify an existing rule
␈↓ α←␈↓	++** ␈↓α1␈↓	

␈↓ α←␈↓	The new rule will be called RULE383
␈↓ α←␈↓	 If:    1 -


␈↓"β␈↓ α←␈↓␈↓ β?The␈α↔expert␈α_has␈α↔tracked␈α↔down␈α_one␈α↔problem: ␈α_The␈α↔performance
␈↓ α←␈↓program␈α∩should␈α∩have␈α∩been␈α∩able␈α∩to␈α∩deduce␈α∩the␈α∩probable␈α∩category␈α∩of␈α∩the
␈↓ α←␈↓␈↓58    EXPLANATION␈↓ 
∪3-10␈↓

␈↓"β␈↓ α←␈↓organism.␈α∂ ␈↓¬TEIRESIAS␈↓␈α∂is␈α∞now␈α∂ready␈α∂to␈α∞accept␈α∂a␈α∂new␈α∞rule␈α∂to␈α∂fix␈α∂the␈α∞problem.
␈↓ α←␈↓We're␈αgoing␈αto␈αstop␈αat␈αthis␈αpoint␈αand␈αstart␈αagain␈αin␈αchapter␈α5␈αin␈αorder␈αto␈αsee
␈↓ α←␈↓how the acquisition process proceeds.

␈↓"β␈↓ α←␈↓␈↓α3-11    SUMMARY␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Three␈α
different␈α
means␈αof␈α
generating␈α
explanations␈α
have␈αbeen␈α
explored
␈↓ α←␈↓in␈αthis␈αchapter.␈α The␈αfirst␈αtechnique,␈αwhich␈αexplores␈αbranches␈αof␈αthe␈αtree␈αnot
␈↓ α←␈↓yet␈α∞traversed␈α∞by␈α∞the␈α∞consultation␈α∞program,␈α∞is␈α∞used␈α∞in␈α∞producing␈α∞answers␈α
to
␈↓ α←␈↓``how''␈αquestions␈α(e.g.,␈α
␈↓↓How␈αwill␈αyou␈α
determine␈αthe␈αidentity␈αof␈α
ORGANISM-1?␈↓).
␈↓ α←␈↓These␈α⊗are␈α⊗produced␈α⊗by␈α↔having␈α⊗the␈α⊗explanation␈α⊗program␈α↔simulate␈α⊗the
␈↓ α←␈↓operation␈α⊃of␈α⊃the␈α⊃consultation␈α⊂program␈α⊃via␈α⊃special-purpose␈α⊃software.␈α⊃ It␈α⊂is
␈↓ α←␈↓thus a hand-crafted solution.
␈↓"β␈↓ α←␈↓␈↓ β?More␈α∪general␈α∀is␈α∪the␈α∪use␈α∀of␈α∪the␈α∪goal-tree␈α∀concept␈α∪as␈α∪a␈α∀basis␈α∪for
␈↓ α←␈↓explanation.␈α⊃ Since␈α⊃the␈α⊃notion␈α∩of␈α⊃a␈α⊃goal␈α⊃tree␈α∩models␈α⊃a␈α⊃large␈α⊃part␈α∩of␈α⊃the
␈↓ α←␈↓control␈α
structure,␈α
it␈α
provides␈α
a␈α
single,␈α
uncomplicated␈α
model␈α
for␈α
much␈α
of␈αthe
␈↓ α←␈↓performance␈α∃program's␈α∀behavior.␈α∃ As␈α∀a␈α∃result,␈α∀a␈α∃simple␈α∃formalism␈α∀that
␈↓ α←␈↓equates␈α⊂``why''␈α⊂and␈α⊂``how''␈α⊂with␈α∂tree␈α⊂traversal␈α⊂offers␈α⊂a␈α⊂reasonably␈α∂powerful
␈↓ α←␈↓and comprehensive explanation capability.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩most␈α⊃general␈α∩technique␈α∩involves␈α⊃having␈α∩directly␈α∩examine␈α⊃the
␈↓ α←␈↓rules␈α⊃in␈α⊂the␈α⊃knowledge␈α⊂base,␈α⊃as␈α⊂in␈α⊃the␈α⊂use␈α⊃of␈α⊂the␈α⊃templates␈α⊃to␈α⊂determine
␈↓ α←␈↓whether␈αa␈αpremise␈αclause␈αhas␈αalready␈αbeen␈αestablished␈αor␈αis␈αstill␈αuntested.␈α In
␈↓ α←␈↓doing␈α∞this,␈α∞␈↓¬TEIRESIAS␈↓'s␈α∞explanation␈α∞facility␈α∞examines␈α∞and␈α∞interprets␈α∞the␈α
same
␈↓ α←␈↓piece␈α∀of␈α∀code␈α∀that␈α∀the␈α∪performance␈α∀program␈α∀is␈α∀about␈α∀to␈α∀execute.␈α∪ The
␈↓ α←␈↓resulting␈αexplanation␈αis␈αthus␈αconstructed␈αwith␈αreference␈αto␈αthe␈αcontent␈αof␈αthe
␈↓ α←␈↓rule,␈α
and␈α
this␈α
referral␈α
is␈αguided␈α
by␈α
information␈α
(the␈α
templates)␈α
contained␈αin
␈↓ α←␈↓the rule components themselves.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α
have␈α
also␈α
seen␈α
two␈αdistinct␈α
uses␈α
for␈α
the␈α
explanations␈αthat␈α
␈↓¬TEIRESIAS␈↓
␈↓ α←␈↓can␈α≤generate.␈α≥ They␈α≤can␈α≤help␈α≥make␈α≤a␈α≤performance␈α≥program␈α≤more
␈↓ α←␈↓comprehensible␈α∩by␈α∩displaying␈α∪the␈α∩reasoning␈α∩it␈α∪employed␈α∩and␈α∩can␈α∪aid␈α∩in
␈↓ α←␈↓uncovering␈α⊂shortcomings␈α⊂in␈α∂the␈α⊂knowledge␈α⊂base.␈α∂ The␈α⊂next␈α⊂three␈α∂chapters
␈↓ α←␈↓follow␈α⊗up␈α↔on␈α⊗this␈α⊗second␈α↔theme␈α⊗and␈α⊗show␈α↔how␈α⊗to␈α⊗rectify␈α↔the␈α⊗errors
␈↓ α←␈↓discovered.
␈↓ α←␈↓␈↓␈↓ 
⊃    59␈↓




␈↓"β␈↓ α←␈↓␈↓ ε∀␈↓αChapter 4



␈↓"β␈↓ α←␈↓α␈↓ βε␈↓λKNOWLEDGE ACQUISITION: 
␈↓ α←␈↓ ¬"␈↓λOVERVIEW␈↓α


␈↓"β␈↓ α←␈↓α␈↓ ¬Ra global perspective









␈↓"β␈↓ α←␈↓␈↓ ¬[The essence of knowledge is, having it, to apply it;
␈↓"β␈↓ α←␈↓␈↓ εVnot having it, to confess your ignorance.
␈↓"β␈↓ α←␈↓␈↓ 	SConfucius

␈↓"β␈↓ α←␈↓␈↓α4-1    INTRODUCTION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Before␈α∞describing␈α∞the␈α∞range␈α∞of␈α∞knowledge␈α∞acquisition␈α∞capabilities␈α
in
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓,␈αthere␈αare␈αa␈αfew␈αintroductory␈αcomments␈αthat␈αwill␈αhelp␈αto␈αestablish␈αa
␈↓ α←␈↓global␈α_perspective␈α_on␈α_what␈α_follows.␈α_ First,␈α_since␈α_the␈α→term␈α_``knowledge
␈↓ α←␈↓acquisition''␈α∀has␈α∀been␈α∀used␈α∀previously␈α∪to␈α∀describe␈α∀a␈α∀range␈α∀of␈α∀tasks,␈α∪we
␈↓ α←␈↓characterize our view of it.

␈↓"β␈↓ α←␈↓␈↓α4-2    PERSPECTIVE ON KNOWLEDGE ACQUISITION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Our␈α⊂efforts␈α⊂have␈α⊂focused␈α⊃on␈α⊂the␈α⊂interactive␈α⊂␈↓↓transfer␈↓␈α⊃of␈α⊂knowledge
␈↓ α←␈↓from␈αan␈αexpert␈αto␈αthe␈αprogram.␈α
 The␈αsystem␈αdoes␈αnot␈αattempt␈αto␈α
derive␈αnew
␈↓ α←␈↓rules␈αon␈αits␈αown,␈αbut␈αinstead␈αtries␈αto␈α``listen''␈αas␈αattentively␈αand␈αintelligently␈αas
␈↓ α←␈↓possible,␈α⊂to␈α∂help␈α⊂the␈α⊂expert␈α∂augment␈α⊂(or␈α⊂edit)␈α∂the␈α⊂knowledge␈α⊂base.␈α∂ Other
␈↓ α←␈↓related work viewed in these terms will illustrate several differences.
␈↓"β␈↓ α←␈↓␈↓ β?Winston's␈α↔work␈α⊗[Winston70]␈α↔on␈α⊗learning␈α↔from␈α⊗examples␈α↔is␈α⊗one
␈↓ α←␈↓instance␈α∪of␈α∀a␈α∪concept-formation␈α∀system.␈α∪ While␈α∪what␈α∀follows␈α∪is␈α∀a␈α∪slight
␈↓ α←␈↓oversimplification,␈α⊃we␈α∩might␈α⊃say␈α∩that␈α⊃in␈α⊃our␈α∩approach,␈α⊃the␈α∩expert␈α⊃would
␈↓ α←␈↓explicitly␈α∪describe␈α∪the␈α∪new␈α∪concept␈α∪to␈α∪the␈α∪system␈α∪rather␈α∪than␈α∀point␈α∪out
␈↓ α←␈↓illustrative␈αexamples.␈α He␈α
would␈αthus␈αsay␈α``An␈α
e.coli␈αis␈αa␈αgram␈α
negative␈αrod''
␈↓ α←␈↓instead of choosing examples from which this fact might be inferred.
␈↓ α←␈↓␈↓60    KNOWLEDGE ACQUISITION:  OVERVIEW␈↓ 
#4-2␈↓

␈↓"β␈↓ α←␈↓␈↓ β?Work␈α≡by␈α≡McDermott␈α≡[McDermott74]␈α≡is␈α≡oriented␈α∨toward␈α≡the
␈↓ α←␈↓acquisition␈α
of␈α
new␈α
information␈α
about␈α∞a␈α
model␈α
world␈α
and␈α
the␈α∞integration␈α
of
␈↓ α←␈↓that␈αinformation␈α
into␈αthe␈α
current␈αdatabase.␈αHe␈α
emphasizes␈αthe␈α
use␈αof␈αa␈α
single
␈↓ α←␈↓formalism␈αfor␈αknowledge␈αrepresentation␈αand␈αfocuses␈αon␈αquestions␈αconcerning
␈↓ α←␈↓belief␈α⊂systems␈α⊂and␈α⊂their␈α∂consistency.␈α⊂ Our␈α⊂system␈α⊂provides␈α⊂more␈α∂flexibility
␈↓ α←␈↓with␈α∀respect␈α∀to␈α∪representations␈α∀and␈α∀has␈α∀not␈α∪yet␈α∀attempted␈α∀to␈α∀deal␈α∪with
␈↓ α←␈↓questions␈α
of␈α
consistency.␈α
 In␈α
addition,␈α
his␈α
system␈α
deals␈α
only␈α
with␈α
acquisition
␈↓ α←␈↓of␈α
new␈α
facts␈α
about␈α
the␈α
world,␈αwhere␈α
we␈α
will␈α
be␈α
very␈α
much␈α
concerned␈αwith␈α
the
␈↓ α←␈↓acquisition of new rules of inference as well.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∞report␈α∞by␈α∞Brachman␈α∞[Brachman75]␈α∞outlines␈α∞some␈α∞issues␈α∞involved
␈↓ α←␈↓in␈α
designing␈α
a␈α
system␈α
similar␈α
in␈α
some␈α
respects␈α
to␈α
the␈α
one␈α
described␈αhere.␈α
 The
␈↓ α←␈↓major␈α⊃point␈α∩of␈α⊃overlap␈α⊃lies␈α∩in␈α⊃his␈α⊃observation␈α∩that␈α⊃one␈α⊃might␈α∩teach␈α⊃the
␈↓ α←␈↓system␈αnot␈α
only␈αnew␈α
instances␈αof␈α
the␈αkind␈α
of␈αobjects␈α
it␈αdeals␈α
with␈α(in␈αhis␈α
case,
␈↓ α←␈↓bibliographic␈α
entries),␈α
but␈α
one␈α
can␈α
teach␈α
the␈α
concept␈α
of␈α
the␈α
object␈αitself.␈α
 This
␈↓ α←␈↓has␈α∞some␈α∞similarity␈α∂to␈α∞the␈α∞perspective␈α∞in␈α∂chapter␈α∞2␈α∞that␈α∂suggested␈α∞teaching
␈↓ α←␈↓the system about its own representations.
␈↓"β␈↓ α←␈↓␈↓ β?It␈α
is␈α
also␈α
useful␈α∞to␈α
compare␈α
a␈α
number␈α∞of␈α
systems␈α
with␈α
respect␈α∞to␈α
the
␈↓ α←␈↓amount␈α⊃and␈α⊃character␈α⊃of␈α⊃their␈α⊃interactions␈α⊃with␈α⊃the␈α⊃user.␈α∩ In␈α⊃␈↓¬META-DENDRAL␈↓
␈↓ α←␈↓[Buchanan72],␈αfor␈α
example,␈αuser␈α
interaction␈αis␈α
limited␈αto␈α
adjustment␈αof␈αa␈α
few
␈↓ α←␈↓performance␈α↔parameters;␈α↔the␈α_system␈α↔operates␈α↔largely␈α_independently.␈α↔ In
␈↓ α←␈↓Winston's␈α
system,␈α
the␈α
interaction␈α
is␈αsomewhat␈α
more␈α
extensive␈α
but␈α
is␈αlimited␈α
to
␈↓ α←␈↓presenting␈αexamples␈α
and␈αindicating␈αwhether␈α
they␈αare␈α
instances␈αof␈αthe␈α
concept
␈↓ α←␈↓in␈α∪question--the␈α∩system␈α∪is␈α∩still␈α∪responsible␈α∩for␈α∪forming␈α∩the␈α∪concepts.␈α∩ In
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓,␈α
interaction␈αwith␈α
the␈αsystem␈α
is␈α
extensive,␈αand␈α
the␈αuser␈α
is␈α
the␈αsource
␈↓ α←␈↓of␈α⊂both␈α⊂the␈α⊂new␈α⊂inference␈α⊂rules␈α⊂and␈α⊂new␈α⊂concepts.␈α⊂ This␈α⊂system␈α⊂does␈α⊂not
␈↓ α←␈↓attempt␈αto␈αinfer␈αnew␈αrules;␈αinstead␈αit␈αtries␈αto␈αmake␈αthe␈αtransfer␈αof␈αknowledge
␈↓ α←␈↓as easy as possible.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α
as␈α
will␈α
become␈α
clear,␈α
our␈α
approach␈α
is␈α
best␈α
suited␈α∞to␈α
learning
␈↓ α←␈↓one␈α
or␈αa␈α
few␈αrules␈α
at␈αa␈α
time.␈α Other␈α
approaches␈αcan␈α
easily␈αbe␈α
imagined.␈α In
␈↓ α←␈↓the␈α∪``exemplary␈α∀programming''␈α∪technique␈α∀described␈α∪in␈α∀[Waterman77],␈α∪for
␈↓ α←␈↓example,␈α⊂the␈α⊃system␈α⊂infers␈α⊃an␈α⊂entire␈α⊃set␈α⊂of␈α⊃rules␈α⊂for␈α⊃a␈α⊂particular␈α⊃task␈α⊂by
␈↓ α←␈↓observing␈αhuman␈αperformance␈αand␈αattempting␈αto␈αwrite␈αrules␈αthat␈α
incorporate
␈↓ α←␈↓the appropriate generalization of that performance.

␈↓"β␈↓ α←␈↓␈↓α4-3    KNOWLEDGE ACQUISITION IN CONTEXT␈↓
␈↓"β␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓␈α_is␈α_designed␈α→to␈α_work␈α_with␈α_performance␈α→programs␈α_that
␈↓ α←␈↓accommodate␈αinexact␈αknowledge.␈α
 Such␈αprograms␈αfind␈α
their␈αgreatest␈αutility␈α
in
␈↓ α←␈↓domains␈α∩where␈α∩knowledge␈α∩has␈α∩not␈α∩been␈α∩extensively␈α∩formalized.␈α∪ In␈α∩such
␈↓ α←␈↓domains␈α∞there␈α∞are␈α∂typically␈α∞no␈α∞unifying␈α∞laws␈α∂on␈α∞which␈α∞to␈α∂base␈α∞algorithmic
␈↓ α←␈↓methods;␈α∃instead␈α∃there␈α∃is␈α∃a␈α∃collection␈α∃of␈α∃informal␈α∃knowledge␈α∃based␈α∀on
␈↓ α←␈↓accumulated␈αexperience.␈α As␈αa␈αresult,␈αan␈αexpert␈αspecifying␈αa␈αnew␈αrule␈αin␈αthis
␈↓ α←␈↓domain␈α∂may␈α⊂be␈α∂codifying␈α∂a␈α⊂piece␈α∂of␈α∂knowledge␈α⊂that␈α∂has␈α⊂never␈α∂previously
␈↓ α←␈↓been␈α
isolated␈αand␈α
expressed␈α
as␈αsuch.␈α
 This␈αprocess␈α
of␈α
explicating␈αpreviously
␈↓ α←␈↓informal␈αknowledge␈αis␈αdifficult,␈αand␈αanything␈αwhich␈αcan␈αbe␈αdone␈αto␈αease␈αthe
␈↓ α←␈↓task will prove very useful.
␈↓ α←␈↓␈↓4-3␈↓ ε4KNOWLEDGE ACQUISITION IN CONTEXT    61␈↓

␈↓"β␈↓ α←␈↓␈↓ β?In␈α↔response,␈α⊗we␈α↔have␈α↔emphasized␈α⊗knowledge␈α↔acquisition␈α↔in␈α⊗the
␈↓ α←␈↓context␈α∞of␈α∂shortcomings␈α∞in␈α∂the␈α∞knowledge␈α∂base.␈α∞ To␈α∂illustrate␈α∞the␈α∂utility␈α∞of
␈↓ α←␈↓this approach, consider the difference between asking the expert

␈↓"β␈↓ α←␈↓␈↓ ∧WWhat should I know about bacteremia?

␈↓ α←␈↓and saying to him

␈↓"β␈↓ α←␈↓␈↓ β'Here␈α∀is␈α∪a␈α∀case␈α∀history␈α∪for␈α∀which␈α∪you␈α∀claim␈α∀the␈α∪performance
␈↓ α←␈↓␈↓ β'program␈α
incorrectly␈α
deduced␈α
the␈α
presence␈α
of␈α
pseudomonas.␈α Here
␈↓ α←␈↓␈↓ β'is␈αhow␈αit␈αreached␈αits␈αconclusions,␈αand␈αhere␈αare␈αall␈αthe␈αfacts␈αof␈αthe
␈↓ α←␈↓␈↓ β'case.␈α Now,␈α␈↓↓what␈αis␈αit␈αthat␈αyou␈αknow␈αthat␈αthe␈αperformance␈αprogram
␈↓ α←␈↓↓␈↓ β'doesn't␈↓, which allows you to avoid making that mistake?␈↓↓

␈↓ α←␈↓Consider␈α⊂how␈α⊃much␈α⊂more␈α⊂focused␈α⊃the␈α⊂second␈α⊂question␈α⊃is␈α⊂and␈α⊃how␈α⊂much
␈↓ α←␈↓easier it would be to explicate the relevant knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂focusing␈α⊂provided␈α⊂by␈α⊂the␈α⊃context␈α⊂is␈α⊂also␈α⊂an␈α⊂important␈α⊃aid␈α⊂to
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓.␈α
 In␈αparticular,␈α
it␈αpermits␈α
the␈αsystem␈α
to␈αbuild␈α
up␈αa␈α
set␈αof␈α
␈↓↓expectations␈↓
␈↓ α←␈↓concerning␈α⊃the␈α⊃knowledge␈α⊃to␈α⊃be␈α⊃expressed,␈α⊃facilitating␈α∩knowledge␈α⊃transfer
␈↓ α←␈↓and making possible several useful features illustrated in the next chapter.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∃approach␈α∃should␈α∃be␈α∀distinguished␈α∃from␈α∃the␈α∃one␈α∀commonly
␈↓ α←␈↓described␈α
by␈α∞the␈α
phrase␈α
␈↓↓knowledge␈α∞as␈α
debugging␈↓,␈α
illustrated␈α∞by␈α
the␈α∞work␈α
of
␈↓ α←␈↓[Sussman75]␈α
and␈α
[Goldstein74].␈α∞ That␈α
technique␈α
suggests␈α
that␈α∞an␈α
important
␈↓ α←␈↓part␈α∞of␈α∞problem␈α∞solving␈α∞is␈α∂knowing␈α∞how␈α∞to␈α∞correct␈α∞``nearly␈α∂right''␈α∞solutions.
␈↓ α←␈↓The␈αemphasis␈α
there␈αis␈α
on␈αdeveloping␈αa␈α
taxonomy␈αof␈α
problem-solving␈αerrors
␈↓ α←␈↓and␈αassembling␈α
a␈αstore␈α
of␈αknowledge␈αabout␈α
corresponding␈αrepairs.␈α
 There␈αis
␈↓ α←␈↓some␈αknowledge␈αof␈αthis␈αsort␈αin␈α␈↓¬TEIRESIAS␈↓'s␈αknowledge␈αacquisition␈αroutines,␈αbut
␈↓ α←␈↓it␈α_is␈α_largely␈α_hand-tailored␈α_and␈α_designed␈α_to␈α_deal␈α_with␈α_details␈α_of␈α↔the
␈↓ α←␈↓performance␈α
program's␈α
operation.␈α
 More␈α
of␈α
it,␈α
and␈α
a␈α
more␈α
general␈α
foundation
␈↓ α←␈↓for␈αit,␈αwould␈αadd␈αuseful␈αcapabilities.␈α They␈αwould␈αbe␈αdistinct,␈αhowever,␈αfrom
␈↓ α←␈↓the␈α↔advantages␈α↔illustrated␈α↔here,␈α↔which␈α↔accrue␈α↔from␈α↔the␈α↔explication␈α⊗of
␈↓ α←␈↓knowledge in the context of an error.

␈↓"β␈↓ α←␈↓␈↓α4-4    KNOWLEDGE BASE MANAGEMENT␈↓
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∩third␈α⊃aspect␈α∩of␈α∩the␈α⊃approach␈α∩in␈α∩␈↓¬TEIRESIAS␈↓␈α⊃lies␈α∩in␈α∩viewing␈α⊃some
␈↓ α←␈↓elements␈α
of␈α∞knowledge␈α
base␈α
management␈α∞in␈α
terms␈α
of␈α∞database␈α
management.
␈↓ α←␈↓That␈α∞is,␈α∞part␈α∞of␈α∞the␈α∞task␈α∞looks␈α∞like␈α∞the␈α∞sort␈α∞of␈α∂data-structure␈α∞manipulation
␈↓ α←␈↓that␈α∞has␈α
been␈α∞the␈α
focus␈α∞of␈α
previous␈α∞work␈α
on␈α∞databases␈α∞(e.g.,␈α
[Sandewall75],
␈↓ α←␈↓[McLeod76],␈α
[Johnson75]).␈α But␈α
our␈αwork␈α
differs␈α
in␈αseveral␈α
respects.␈α For␈α
one
␈↓ α←␈↓thing,␈α∩it␈α∪involves␈α∩constructing␈α∪an␈α∩integral␈α∩part␈α∪of␈α∩the␈α∪high␈α∩performance
␈↓ α←␈↓program,␈αrather␈αthan␈αsimply␈αa␈αdatabase␈αfrom␈αwhich␈αinformation␈αis␈αretrieved.
␈↓ α←␈↓In␈α
most␈αstandard␈α
database␈α
tasks,␈αthe␈α
retrieval␈α
or␈αstorage␈α
of␈α
information␈αis␈α
the
␈↓ α←␈↓ultimate␈α
aim.␈α∞ In␈α
␈↓¬TEIRESIAS␈↓,␈α∞we␈α
are␈α
assembling␈α∞large␈α
amounts␈α∞of␈α
information
␈↓ α←␈↓that␈α⊂will␈α⊂be␈α⊂used␈α⊂by␈α⊂the␈α⊃performance␈α⊂program␈α⊂and␈α⊂that␈α⊂will␈α⊂enable␈α⊃it␈α⊂to
␈↓ α←␈↓reason␈α∞about␈α∞the␈α∞domain.␈α∞ Where␈α∞database␈α∞work␈α∞concentrates␈α∞solely␈α∞on␈α∞the
␈↓ α←␈↓␈↓62    KNOWLEDGE ACQUISITION:  OVERVIEW␈↓ 
#4-4␈↓

␈↓"β␈↓ α←␈↓management␈αof␈αfacts,␈αour␈αsystem␈αis␈αalso␈αconcerned␈αwith␈αthe␈αmanagement␈αof␈αa
␈↓ α←␈↓set of inference rules.
␈↓"β␈↓ α←␈↓␈↓ β?Another␈α∩basic␈α∩difference␈α∪is␈α∩the␈α∩reliance␈α∩on␈α∪meta-level␈α∩knowledge.
␈↓ α←␈↓Examples␈αin␈αlater␈αchapters␈α
will␈αillustrate␈αthat␈αmanagement␈αof␈α
the␈αknowledge
␈↓ α←␈↓base␈α
can␈α
be␈α
founded␈α
on␈α
both␈α
the␈α
system's␈α
access␈α
to␈α
and␈α
``understanding''␈αof␈α
its
␈↓ α←␈↓own representations.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α∪our␈α∪system␈α∪focuses␈α∪primarily␈α∪on␈α∪making␈α∪additions␈α∪to␈α∩the
␈↓ α←␈↓knowledge␈α∃base.␈↓
1␈↓␈α∃Concentrating␈α∃on␈α∃this␈α∃single␈α∃operation␈α∃will␈α⊗focus␈α∃the
␈↓ α←␈↓discussion yet still cover most of the interesting problems.

␈↓"β␈↓ α←␈↓␈↓α4-5    SYSTEM DIAGRAM␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Fig.␈α↔2-3␈α⊗offered␈α↔one␈α↔view␈α⊗of␈α↔␈↓¬TEIRESIAS␈↓␈α↔and␈α⊗its␈α↔relation␈α↔to␈α⊗the
␈↓ α←␈↓performance␈α∃program.␈α∀ ``Zooming␈α∃in''␈α∀on␈α∃the␈α∀knowledge␈α∃acquisition␈α∀box
␈↓ α←␈↓shown␈α∂in␈α∂that␈α∂figure␈α∂produces␈α∂the␈α∂view␈α∂presented␈α∂in␈α∂Fig.␈α∂4-1,␈α∂which␈α∂is␈α∂a
␈↓ α←␈↓complete␈α∂overview␈α∂of␈α∂the␈α∂capabilities␈α⊂described␈α∂in␈α∂chapters␈α∂5␈α∂and␈α⊂6.␈α∂ The
␈↓ α←␈↓figure␈α
shows␈αboth␈α
the␈αprocesses␈α
(standard␈α
boxes)␈αand␈α
different␈αrepositories␈α
of
␈↓ α←␈↓knowledge␈α∂(double-walled␈α∂boxes)␈α⊂divided␈α∂up␈α∂according␈α∂to␈α⊂their␈α∂conceptual
␈↓ α←␈↓appearance␈α⊃(rather␈α⊃than␈α⊃their␈α∩physical␈α⊃structure).␈α⊃ Both␈α⊃the␈α∩``new␈α⊃schema
␈↓ α←␈↓acquisition''␈α
and␈α∞``new␈α
instance␈α
acquisition''␈α∞are␈α
accomplished␈α
with␈α∞the␈α
same
␈↓ α←␈↓body␈α
of␈α
code␈α
(and␈α
are␈α
described␈αin␈α
chapter␈α
6).␈α
 Other␈α
paths␈α
of␈αinformation
␈↓ α←␈↓flow␈αhave␈αbeen␈αomitted␈αfrom␈αthis␈αdiagram␈αfor␈αthe␈αsake␈αof␈αclarity.␈α They␈αwill
␈↓ α←␈↓all be described in the next two chapters.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∀final␈α∀introductory␈α∀point␈α∀to␈α∀be␈α∀made␈α∀is␈α∀the␈α∀mixed-initiative
␈↓ α←␈↓nature␈αof␈αthe␈α␈↓¬TEIRESIAS␈↓-expert␈αinteraction.␈α
 In␈αgeneral,␈αthe␈αuser␈αindicates␈α
what
␈↓ α←␈↓he␈αwants␈αto␈αdo.␈α ␈↓¬TEIRESIAS␈↓␈αthen␈αtakes␈αover␈αand␈αstructures␈αthe␈αtask,␈αrequesting
␈↓ α←␈↓information␈α∀from␈α∀the␈α∀user␈α∃and␈α∀getting␈α∀his␈α∀approval␈α∀before␈α∃taking␈α∀any
␈↓ α←␈↓important␈αsteps.␈α Numerous␈αexamples␈αof␈αthis␈αwill␈αbe␈αencountered␈αin␈αthe␈αnext
␈↓ α←␈↓two chapters.














␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[1]␈α∀Since␈α∀deletion␈α∀of␈α∀structures␈α∃can␈α∀often␈α∀be␈α∀viewed␈α∀as␈α∀the␈α∃inverse␈α∀of
␈↓ α←␈↓addition,␈α⊃and␈α∩modification␈α⊃viewed␈α∩as␈α⊃deletion␈α∩followed␈α⊃by␈α∩addition,␈α⊃little
␈↓ α←␈↓generality is lost.
␈↓"β␈↓ α←␈↓␈↓4-5␈↓ λ5SYSTEM DIAGRAM    63␈↓


␈↓"␈↓ α←␈↓∧                      KNOWLEDGE ACQUISITION
␈↓"␈↓ α←␈↓∧            ⊂ααααααααααααααααααααααααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            ~ ⊂παααααααααααααααπ⊃                   ~
␈↓"␈↓ α←␈↓∧            ~ ~~2              ~~α α α α α ⊃        ~
␈↓"␈↓ α←␈↓∧            ~ ~~ schema-schema ~~          ↓        ~
␈↓"␈↓ α←␈↓∧            ~ %∀ααααααααααααααα∀$ ⊂αααααααααααααα⊃  ~
␈↓"␈↓ α←␈↓∧            ~         ⊂ α α α α α ~ new schema   ~  ~
␈↓"␈↓ α←␈↓∧            ~         ↓           ~ acquisition  ~  ~
␈↓"␈↓ α←␈↓∧            ~ ⊂παααααααααααααααπ⊃ %αααααααααααααα$  ~
␈↓"␈↓ α←␈↓∧            ~ ~~1              ~~α α α α α ⊃        ~
␈↓"␈↓ α←␈↓∧            ~ ~~    schemata   ~~          ↓        ~
␈↓"␈↓ α←␈↓∧ KNOWLEDGE  ~ %∀ααααααααααααααα∀$ ⊂αααααααααααααα⊃  ~
␈↓"␈↓ α←␈↓∧   BASE     ~                     ~ new instance ~  ~
␈↓"␈↓ α←␈↓∧⊂πααααααπ⊃  ~                ⊂ α α~ acquisition  ~  ~← α EXPERT
␈↓"␈↓ α←␈↓∧~~0     ~~←α~α α α α α α α α $    %αααααααααααααα$  ~[dialog]
␈↓"␈↓ α←␈↓∧~~ facts~~  ~    [knowledge                         ~
␈↓"␈↓ α←␈↓∧~~ -----~~  ~     transfer]                         ~
␈↓"␈↓ α←␈↓∧~~ rules~~←α~α α α α α α α α ⊃    ⊂αααααααααααααα⊃  ~
␈↓"␈↓ α←␈↓∧~~      ~~  ~                % α α~     rule     ~  ~
␈↓"␈↓ α←␈↓∧%∀αααααα∀$  ~                     ~  acquisition ~  ~
␈↓"␈↓ α←␈↓∧            ~                     %αααααααααααααα$  ~
␈↓"␈↓ α←␈↓∧    ~       ~  ⊂παααααααααααααπ⊃          ↑         ~
␈↓"␈↓ α←␈↓∧            ~  ~~1            ~~                    ~
␈↓"␈↓ α←␈↓∧    % α α α ~α→~~     rule    ~~α α α α α $         ~
␈↓"␈↓ α←␈↓∧ [concept   ~  ~~    models   ~~    [model-directed ~
␈↓"␈↓ α←␈↓∧  formation]~  %∀ααααααααααααα∀$     understanding] ~
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            %ααααααααααααααααααααααααααααααααααααααα$


␈↓"
␈↓"β␈↓ α←␈↓∧␈↓ β'␈↓αFig.␈α∀4-1.    Close-up␈α∀view␈α∀of␈α∃knowledge␈α∀acquisition.  ␈↓Double-
␈↓ α←␈↓␈↓ β'walled␈αboxes␈αwith␈αnumbers␈αin␈αthe␈αupper␈αleft-hand␈αcorner␈αcontain
␈↓ α←␈↓␈↓ β'various␈α∩sorts␈α∪of␈α∩knowledge;␈α∪the␈α∩number␈α∪indicates␈α∩the␈α∪level␈α∩of
␈↓ α←␈↓␈↓ β'knowledge␈α↔they␈α_contain.␈α↔ The␈α↔original␈α_performance␈α↔program
␈↓ α←␈↓␈↓ β'knowledge␈α∂base␈α∂is␈α⊂at␈α∂the␈α∂left␈α⊂with␈α∂a␈α∂``␈↓∧0␈↓,''␈α⊂indicating␈α∂object-level
␈↓ α←␈↓␈↓ β'knowledge.␈α∂ Standard␈α∂boxes␈α∂without␈α∂numbers␈α∂indicate␈α⊂the␈α∂three
␈↓ α←␈↓␈↓ β'knowledge␈α~acquisition␈α~processes␈α~available␈α~for␈α~educating␈α→the
␈↓ α←␈↓␈↓ β'system.␈α Dashed␈αarrows␈αindicate␈αinformation␈αflow␈αand␈αare␈αlabeled
␈↓ α←␈↓␈↓ β'(in brackets) when that information flow has a familiar name.
␈↓ α←␈↓␈↓␈↓ 
⊃    65␈↓




␈↓"β␈↓ α←␈↓␈↓ ε∃␈↓αChapter 5



␈↓"β␈↓ α←␈↓α␈↓ β␈↓λKNOWLEDGE ACQUISITION I␈↓α


␈↓"β␈↓ α←␈↓α␈↓ ¬↔learning new inference rules









␈↓"β␈↓ α←␈↓␈↓ ¬GKnowledge␈αis␈αthat␈αsmall␈αpart␈αof␈αignorance␈αthat␈αwe
␈↓ α←␈↓␈↓ ¬Garrange and classify.
␈↓"β␈↓ α←␈↓␈↓ 	∀Ambrose Bierce

␈↓"β␈↓ α←␈↓␈↓α5-1    INTRODUCTION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?When␈αlast␈α
we␈αleft␈α
him␈αin␈αchapter␈α
3,␈αthe␈α
expert␈αhad␈αnoticed␈α
something
␈↓ α←␈↓wrong␈α∀with␈α∀the␈α∀identities␈α∀deduced␈α∀and␈α∀had␈α∀tracked␈α∀down␈α∀the␈α∀problem
␈↓ α←␈↓himself,␈α
using␈α
the␈α
explanation␈α
facilities␈α
in␈α
␈↓¬TEIRESIAS␈↓.␈α
 In␈α
this␈α
chapter,␈αthe␈α
same
␈↓ α←␈↓example␈α∪is␈α∀explored␈α∪to␈α∪show␈α∀how␈α∪the␈α∪system␈α∀can␈α∪direct␈α∀the␈α∪debugging
␈↓ α←␈↓process.␈α Then,␈αin␈αorder␈α
to␈αpresent␈αthe␈αentire␈α
interaction␈αat␈αonce,␈αthe␈α
example
␈↓ α←␈↓continues␈αthrough␈αthe␈αacquisition␈αof␈αa␈αnew␈αrule␈αthat␈αfixes␈αthe␈αproblem.␈α The
␈↓ α←␈↓remainder␈α
of␈α∞the␈α
chapter␈α∞presents␈α
a␈α
global␈α∞overview␈α
of␈α∞the␈α
ideas␈α∞used␈α
that
␈↓ α←␈↓make␈α⊂possible␈α∂the␈α⊂system's␈α⊂performance␈α∂and␈α⊂reviews␈α⊂the␈α∂trace␈α⊂in␈α⊂detail␈α∂to
␈↓ α←␈↓explain the basis for some of the capabilities displayed.
␈↓ α←␈↓␈↓66    KNOWLEDGE ACQUISITION I␈↓ 
#5-1␈↓

␈↓"β␈↓ α←␈↓␈↓α5-2    TRACE OF SYSTEM PERFORMANCE:  DEBUGGING EXAMPLE
␈↓ α←␈↓α␈↓ β3CONTINUED␈↓

␈↓ α←␈↓␈↓ β?It␈α
should␈αbe␈α
emphasized␈αthat␈α
this␈αexample␈α
has␈αbeen␈α
chosen␈αbecause␈α
it
␈↓ α←␈↓displays,␈α∞in␈α
a␈α∞single␈α
pass,␈α∞a␈α
large␈α∞number␈α
of␈α∞the␈α
features␈α∞of␈α
the␈α∞system.␈α
 In
␈↓ α←␈↓particular,␈α
it␈α
shows␈α
initial␈α
failure␈αof␈α
the␈α
primary␈α
natural␈α
language␈αhandling
␈↓ α←␈↓routines␈α⊂and␈α⊂allows␈α⊂the␈α⊂demonstration␈α⊂of␈α⊂many␈α⊂of␈α⊂the␈α⊂secondary␈α⊂features.
␈↓ α←␈↓The␈α→interaction␈α→shown␈α→below␈α→is␈α→thus␈α→longer␈α→than␈α→most.␈α→ As␈α→before,
␈↓ α←␈↓annotations are in italics, and user responses in boldface capitals.

␈↓ α←␈↓	***************************************************************

␈↓ α←␈↓	My therapy recommendation will be based on the following
␈↓ α←␈↓	probable infection(s) and potential causative organism(s):

␈↓ α←␈↓	INFECTION-1 is PRIMARY-BACTEREMIA
␈↓ α←␈↓	    <Item 1>  E.COLI [ORGANISM-1]
␈↓ α←␈↓	    <Item 2>  PSEUDOMONAS-AERUGINOSA [ORGANISM-1]

␈↓ α←␈↓	Are all these identities correct?
␈↓ α←␈↓	++**␈↓α N␈↓	

␈↓ α←␈↓↓␈↓ βW{As␈α∩before,␈α⊃the␈α∩expert␈α⊃disagrees␈α∩with␈α⊃the␈α∩results␈α⊃and␈α∩indicates␈α⊃to
␈↓ α←␈↓↓␈↓ βW␈↓¬TEIRESIAS␈↓↓ the identities that he thinks should have resulted.}

␈↓ α←␈↓	Are there any identities missing?
␈↓ α←␈↓	++**␈↓α Y␈↓	
␈↓ α←␈↓	Which one(s)?
␈↓ α←␈↓	++**␈↓α KLEBSIELLA-PNEUMONIAE   PROTEUS-NON-MIRABILIS␈↓	
␈↓ α←␈↓	Are there any identities that shouldn't have appeared?
␈↓ α←␈↓	++**␈↓α Y␈↓	
␈↓ α←␈↓	Which one(s)?
␈↓ α←␈↓	++**␈↓α PSEUDOMONAS-AERUGINOSA␈↓	

␈↓ α←␈↓	Do you want to
␈↓ α←␈↓	      1] track down the problem yourself, or
␈↓ α←␈↓	      2] allow me to guide the debugging, or
␈↓ α←␈↓	      3] ignore it for now and fix it another time?
␈↓ α←␈↓	Type 1, 2 or 3
␈↓ α←␈↓	++**␈↓α 2␈↓	

␈↓ α←␈↓↓␈↓ βW{Recall that in chapter 3 the expert followed option #1.}

␈↓ α←␈↓	We'll start with the PSEUDOMONAS-AERUGINOSA first.

␈↓ α←␈↓↓␈↓ βW{This␈α⊂time,␈α⊃␈↓¬TEIRESIAS␈↓↓␈α⊂will␈α⊃guide␈α⊂the␈α⊂process.␈α⊃ Again␈α⊂it␈α⊃applies␈α⊂the
␈↓ α←␈↓↓␈↓ βWheuristic of starting with an identity that appeared incorrectly.}
␈↓"β
␈↓ α←␈↓	The following rules were used in deducing that the identity of
␈↓ α←␈↓␈↓5-2␈↓ εoDEBUGGING EXAMPLE CONTINUED    67␈↓

␈↓"β␈↓ α←␈↓	ORGANISM-1 is pseudomonas-aeruginosa

␈↓ α←␈↓	RULE184
␈↓ α←␈↓	-------
␈↓ α←␈↓	Since [1.1] the category of ORGANISM-1 is not known
␈↓ α←␈↓	      [1.2] the gram stain of ORGANISM-1 is gramneg
␈↓ α←␈↓	      [1.3] the morphology of ORGANISM-1 is rod
␈↓ α←␈↓	      [1.4] the aerobicity of ORGANISM-1 is facultative
␈↓ α←␈↓	 There is weakly suggestive evidence (.3) that the identity of
␈↓ α←␈↓	 ORGANISM-1 is pseudomonas-aeruginosa

␈↓ α←␈↓	Is this rule correct?
␈↓ α←␈↓	++**␈↓α Y␈↓	
␈↓ α←␈↓	Should its premise have failed for this case?
␈↓ α←␈↓	++**␈↓α Y␈↓	
␈↓ α←␈↓	Which clause of the premise should have been false? [give #]
␈↓ α←␈↓	++**␈↓α 1␈↓	


␈↓ α←␈↓	No conclusions were reached concerning the category of
␈↓ α←␈↓	ORGANISM-1.

␈↓ α←␈↓	Should it have been possible to conclude a category?
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓	Which one?
␈↓ α←␈↓	++**␈↓α ENTEROBACTERIACEAE␈↓	

␈↓ α←␈↓	There were no rules successfully invoked in the attempt to
␈↓ α←␈↓	deduce that the category of ORGANISM-1 is enterobacteriaceae.

␈↓ α←␈↓	The following (potentially applicable) rules were NOT executed
␈↓ α←␈↓	because the starred [**] clause failed:

␈↓ α←␈↓	RULE116
␈↓ α←␈↓	-------
␈↓ α←␈↓	If    1) the identity of ORGANISM-1 is not known
␈↓ α←␈↓	   ** 2) the gram stain of ORGANISM-1 is not known **
␈↓ α←␈↓	      3) the morphology of ORGANISM-1 is not known
␈↓ α←␈↓	      4) the site of CULTURE-1 is csf
␈↓ α←␈↓	      5) the infection is meningitis
␈↓ α←␈↓	      6) the age (in years) of Albert Hall is less than or
␈↓ α←␈↓	         equal to .17
␈↓ α←␈↓	then  There is weakly suggestive evidence (.33) that the
␈↓ α←␈↓	      category of ORGANISM-1 is enterobacteriaceae
␈↓ α←␈↓␈↓68    KNOWLEDGE ACQUISITION I␈↓ 
#5-2␈↓

␈↓ α←␈↓	RULE050
␈↓ α←␈↓	-------
␈↓ α←␈↓	If    1) the morphology of ORGANISM-1 is rod
␈↓ α←␈↓	      2) the gram stain of ORGANISM-1 is gramneg
␈↓ α←␈↓	      3) the aerobicity of ORGANISM-1 is facultative
␈↓ α←␈↓	   ** 4) the infection with ORGANISM-1 was acquired while the
␈↓ α←␈↓	         patient was hospitalized **
␈↓ α←␈↓	then  There is suggestive evidence (.7) that the category of
␈↓ α←␈↓	      ORGANISM-1 is enterobacteriaceae

␈↓ α←␈↓	Are all these rules correct?
␈↓ α←␈↓	++**␈↓α Y␈↓	
␈↓ α←␈↓	Should any of them have been successfully invoked?
␈↓ α←␈↓	++**␈↓α N␈↓	

␈↓ α←␈↓↓␈↓ βW{This␈α⊂would␈α∂be␈α⊂the␈α∂case␈α⊂if␈α⊂one␈α∂of␈α⊂the␈α∂starred␈α⊂premise␈α⊂clauses␈α∂was
␈↓ α←␈↓↓␈↓ βWfalse␈α∂because␈α∂it␈α⊂had␈α∂been␈α∂``stepped␈α∂on''␈α⊂by␈α∂yet␈α∂another␈α⊂rule␈α∂further
␈↓ α←␈↓↓␈↓ βWback in the chain.}

␈↓ α←␈↓	Is there a rule missing?
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓	Ok, then I need a rule that would allow me to deduce that the
␈↓ α←␈↓	category of ORGANISM-1 is enterobacteriaceae.

␈↓ α←␈↓	Would you like to give me such a rule now?
␈↓ α←␈↓	++**␈↓α Y␈↓	
␈↓ α←␈↓	Good...

␈↓ α←␈↓↓␈↓ βW{To␈αreview␈αbriefly: ␈αThe␈αexpert␈αis␈αtracking␈αdown␈αthe␈αreason␈αwhy␈α␈↓¬MYCIN␈↓↓
␈↓ α←␈↓↓␈↓ βWincorrectly␈α∞concluded␈α∞that␈α∞the␈α∞identity␈α∂was␈α∞pseudomonas-aeruginosa.
␈↓ α←␈↓↓␈↓ βWShown below is the sequence of rules which were involved:
␈↓ α←␈↓¬IDENTITY
␈↓ α←␈↓¬GRAM STAIN **
␈↓ α←␈↓¬MORPHOLOGY
␈↓ α←␈↓¬SITE
␈↓ α←␈↓¬INFECTION
␈↓ α←␈↓¬AGE         --- RULE116 -> \
␈↓ α←␈↓¬                            >→ CATEGORY
␈↓ α←␈↓¬MORPHOLOGY  --- RULE050 -> /   GRAM STAIN   --- RULE184 ---> IDENTITY = PSEUDOMONAS-
␈↓ α←␈↓¬GRAM STAIN                     AEROBICITY                               AERUGINOSA
␈↓ α←␈↓¬AEROBICITY                     MORPHOLOGY
␈↓ α←␈↓¬HOSPITALIZED **
␈↓ α←␈↓↓␈↓ βWPseudomonas-aeruginosa␈α∨appeared␈α∨because␈α∨RULE184␈α∨executed
␈↓ α←␈↓↓␈↓ βWsuccessfully.
␈↓ α←␈↓↓␈↓ βWThis occurred in part because category was unknown.
␈↓ α←␈↓↓␈↓ βWCategory␈α
was␈αunknown␈α
because␈α
all␈αrules␈α
to␈α
conclude␈αit␈α
(50␈α
and␈α116)
␈↓ α←␈↓↓␈↓ βWfailed.
␈↓ α←␈↓↓␈↓ βWSince␈αthe␈α
expert␈αbelieves␈α
all␈αexisting␈αrules␈α
are␈αcorrect,␈α
the␈αknowledge
␈↓ α←␈↓␈↓5-2␈↓ εoDEBUGGING EXAMPLE CONTINUED    69␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βWbase␈α⊃must␈α⊃be␈α⊃missing␈α⊃a␈α⊂rule␈α⊃that␈α⊃will␈α⊃conclude␈α⊃about␈α⊃category;␈α⊂in
␈↓ α←␈↓↓␈↓ βWparticular,␈α⊃that␈α⊃the␈α⊃category␈α⊂is␈α⊃enterobacteriaceae.␈α⊃ This␈α⊃will␈α⊂cause
␈↓ α←␈↓↓␈↓ βWcategory␈α∂to␈α∂be␈α∂known,␈α∂inhibiting␈α∂rule␈α∂184,␈α∂fixing␈α∂at␈α∂least␈α⊂this␈α∂first
␈↓ α←␈↓↓␈↓ βWproblem, the incorrect conclusion about pseudomonas-aeruginosa.
␈↓ α←␈↓↓␈↓ βWNote␈α∞how␈α∂sharply␈α∞focused␈α∂the␈α∞issue␈α∂has␈α∞become,␈α∂for␈α∞both␈α∂the␈α∞expert
␈↓ α←␈↓↓␈↓ βWand␈α_␈↓¬TEIRESIAS␈↓↓,␈α_as␈α_a␈α_result␈α_of␈α_the␈α_large␈α_amount␈α→of␈α_contextual
␈↓ α←␈↓↓␈↓ βWinformation␈α∪available.␈α∪ The␈α∪doctor␈α∪is␈α∪not␈α∪being␈α∪asked␈α∪to␈α∪``discuss
␈↓ α←␈↓↓␈↓ βWbacteremia''; instead, the system can at this point say
␈↓ α←␈↓∧␈↓ ∧∨I␈α∃need␈α∃a␈α∃rule␈α∃that␈α∀would␈α∃allow␈α∃me␈α∃to␈α∃deduce␈α∀that
␈↓ α←␈↓∧␈↓ ∧∨the category of ORGANISM-1 is enterobacteriaceae.
␈↓ α←␈↓↓␈↓ βWand␈αit␈α
must␈αbe␈α
a␈αrule␈αwhich␈α
is␈αinvocable␈α
in␈αthe␈α
context␈αof␈αthis␈α
patient.
␈↓ α←␈↓↓␈↓ βWThis␈α
focus␈α
makes␈α
it␈α
easier␈α
for␈α
the␈α
expert␈α
to␈α
specify␈α
a␈α
rule␈α∞that␈α
may
␈↓ α←␈↓↓␈↓ βWnever␈α
have␈α
been␈α
explicitly␈α
stated␈α
or␈α
recognized␈α
as␈α
such.␈α
This␈α
can␈αbe␈α
an
␈↓ α←␈↓↓␈↓ βWimportant␈α∞advantage␈α
in␈α∞any␈α∞domain␈α
where␈α∞the␈α
knowledge␈α∞is␈α∞not␈α
yet
␈↓ α←␈↓↓␈↓ βWfully formalized.
␈↓ α←␈↓↓␈↓ βWThe␈α
contextual␈α
information␈α
is␈α
useful␈α
for␈α
␈↓¬TEIRESIAS␈↓↓␈α
as␈α
well.␈α
 It␈α
provides
␈↓ α←␈↓↓␈↓ βWa␈αnumber␈αof␈αexpectations␈αabout␈αthe␈α
content␈αof␈αthe␈αrule␈αthat␈αare␈α
useful
␈↓ α←␈↓↓␈↓ βWboth␈αin␈α
understanding␈αwhat␈α
the␈αexpert␈α
types␈αin␈α
and␈αin␈α
checking␈αthe
␈↓ α←␈↓↓␈↓ βWnew␈αrule␈αto␈αmake␈αsure␈αthat␈αit␈αdoes␈αindeed␈αsolve␈αthe␈αproblem␈αat␈αhand.
␈↓ α←␈↓↓␈↓ βWDetails of this process are explained below.
␈↓ α←␈↓↓␈↓ βWThe trace continues with the expert typing in the rule....}

␈↓ α←␈↓	The new rule will be called RULE383
␈↓ α←␈↓	 If:    1 -  ␈↓αTHE PATIENT'S INFECTION IS PRIMARY-␈↓	
␈↓ α←␈↓	             ␈↓αBACTEREMIA␈↓	
␈↓ α←␈↓	   and  2 -  ␈↓αTHE SITE IS ONE OF THE STERILE SITES␈↓	
␈↓ α←␈↓	   and  3 -
␈↓ α←␈↓	 Then:  1 -  ␈↓αTHERE IS EVIDENCE (.8) THAT THE CATEGORY IS␈↓	
␈↓ α←␈↓	             ␈↓αENTEROBACTERIACEAE␈↓	
␈↓ α←␈↓	   and  2 -

␈↓ α←␈↓	This may take a few moments...

␈↓ α←␈↓↓␈↓ βW{Generally, it takes on the order of 10 to 30 seconds of cpu time.}

␈↓ α←␈↓	This is my understanding of your rule:
␈↓ α←␈↓	RULE383
␈↓ α←␈↓	-------
␈↓ α←␈↓	  If:  1) Primary-bacteremia is an associated infection for
␈↓ α←␈↓	          which therapy should cover, and
␈↓ α←␈↓	       2) The infection with the organism was acquired while
␈↓ α←␈↓	          the patient was hospitalized, and
␈↓ α←␈↓	       3) The culture was taken from a sterile source, and
␈↓ α←␈↓	       4) The organism is not normally found at this site
␈↓ α←␈↓	  Then:  There is strongly suggestive evidence (.8) that the
␈↓ α←␈↓	         category of the organism is enterobacteriaceae

␈↓ α←␈↓↓␈↓ βW{␈↓¬TEIRESIAS␈↓↓'s␈αinitial␈αinterpretation␈αof␈αthe␈αnatural␈αlanguage␈αis␈αincorrect
␈↓ α←␈↓␈↓70    KNOWLEDGE ACQUISITION I␈↓ 
#5-2␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βW(recall␈α
that␈α
the␈αexample␈α
was␈α
chosen␈αin␈α
part␈α
for␈αthis␈α
reason).␈α
 In␈αthe
␈↓ α←␈↓↓␈↓ βWsystem's␈α
first␈αattempt␈α
to␈α
understand␈αthe␈α
rule,␈α
each␈αline␈α
of␈αthe␈α
original
␈↓ α←␈↓↓␈↓ βWtext␈αhas␈αbeen␈αinterpreted␈αas␈αtwo␈αclauses.␈α This␈αresults␈αprimarily␈αfrom
␈↓ α←␈↓↓␈↓ βWthe␈α⊃word-by-word␈α⊂approach␈α⊃to␈α⊂meaning.␈α⊃ For␈α⊂instance,␈α⊃despite␈α⊂the
␈↓ α←␈↓↓␈↓ βWobvious␈α⊂grammatical␈α⊂simplicity␈α⊂of␈α⊃the␈α⊂first␈α⊂line␈α⊂of␈α⊂text,␈α⊃the␈α⊂system
␈↓ α←␈↓↓␈↓ βWused␈α∞the␈α∞word␈α∞``primary-bacteremia''␈α∞as␈α∞the␈α∞basis␈α∞for␈α∞generating␈α
the
␈↓ α←␈↓↓␈↓ βWfirst␈α
clause,␈α
and␈αit␈α
used␈α
the␈α
word␈α``infection''␈α
to␈α
generate␈α
the␈αsecond.
␈↓ α←␈↓↓␈↓ βWIn␈α
the␈αsecond␈α
line␈αof␈α
text␈α
the␈αword␈α
``sterile''␈αwas␈α
responsible␈αfor␈α
clause
␈↓ α←␈↓↓␈↓ βW#3,␈αand␈αthe␈αword␈α``sites''␈αused␈αas␈αthe␈αbasis␈αfor␈αclause␈α#4.␈α The␈αdetails
␈↓ α←␈↓↓␈↓ βWof␈α∂this␈α∂process␈α∂are␈α∂explained␈α∂below,␈α∂where␈α∂it␈α∂will␈α∂become␈α⊂clear␈α∂that
␈↓ α←␈↓↓␈↓ βWwhile␈α
the␈α
translation␈αis␈α
wrong,␈α
it␈α
is␈αnot␈α
unreasonable␈α
given␈αthe␈α
simple
␈↓ α←␈↓↓␈↓ βWnatural language facilities.
␈↓ α←␈↓↓␈↓ βWNor␈αis␈αit␈αunreasonable␈αthat␈α␈↓¬TEIRESIAS␈↓↓␈αhas␈αturned␈αeach␈αline␈αof␈αEnglish
␈↓ α←␈↓↓␈↓ βWinto␈α∂more␈α∂than␈α⊂one␈α∂premise␈α∂clause.␈α∂ The␈α⊂expert␈α∂is␈α∂not␈α⊂restricted␈α∂to
␈↓ α←␈↓↓␈↓ βWtyping␈αthe␈α
English␈αequivalent␈α
of␈αa␈α
single␈αpremise␈α
clause␈αon␈αeach␈α
line.
␈↓ α←␈↓↓␈↓ βWIf␈α→he␈α~were,␈α→rather␈α~than␈α→typing␈α→``␈↓	the␈α~organism␈α→is␈α~a␈α→gram
␈↓ α←␈↓	␈↓ βWnegative␈α⊂aerobic␈α⊂rod␈↓↓,''␈α∂he␈α⊂would␈α⊂have␈α∂to␈α⊂type␈α⊂three␈α⊂lines,␈α∂``␈↓	the
␈↓ α←␈↓	␈↓ βWorganism␈α_is␈α↔gram␈α_negative␈↓↓,''␈α↔``␈↓	it␈α_is␈α↔aerobic␈↓↓,''␈α_``␈↓	it␈α_is␈α↔a
␈↓ α←␈↓	␈↓ βWrod␈↓↓.'' ␈α⊂The␈α⊂cost␈α⊂of␈α⊂this␈α⊂capability,␈α⊂however,␈α⊂is␈α⊂shown␈α⊂here: ␈α⊂It␈α∂may
␈↓ α←␈↓↓␈↓ βWresult in spurious clauses.
␈↓ α←␈↓↓␈↓ βWWe␈α∞will␈α∞see␈α
later␈α∞that␈α∞while␈α
␈↓¬TEIRESIAS␈↓↓'s␈α∞interpretations␈α∞are␈α
incorrect,
␈↓ α←␈↓↓␈↓ βWthey␈α⊂are␈α⊂all␈α∂``reasonable''␈α⊂and␈α⊂``consistent,''␈α∂in␈α⊂light␈α⊂of␈α⊂the␈α∂system's
␈↓ α←␈↓↓␈↓ βWsomewhat␈α∞primitive␈α∂approach␈α∞to␈α∂natural␈α∞language.␈α∂ The␈α∞rest␈α∂of␈α∞the
␈↓ α←␈↓↓␈↓ βWexample␈α∀shows␈α∀that␈α∀even␈α∃when␈α∀the␈α∀initial␈α∀guesses␈α∃are␈α∀incorrect,
␈↓ α←␈↓↓␈↓ βWstraightening things out may conceivably be very easy.}

␈↓ α←␈↓	Okay? (Yes or No)
␈↓ α←␈↓	++**␈↓α N␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈αexpert␈αindicates␈α
that␈αthere␈αis␈α
a␈αproblem␈αwith␈α
the␈αinterpretation
␈↓ α←␈↓↓␈↓ βWand uses the system's rule editor to specify the changes needed.}

␈↓ α←␈↓	Change premise?
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓	What clause should be changed in premise?
␈↓ α←␈↓	(type HELP for list of commands)
␈↓ α←␈↓	++**␈↓α CHANGE 1␈↓	

␈↓ α←␈↓	Suppose (1) were changed to read-
␈↓ α←␈↓	        1 - The infection is primary-bacteremia
␈↓ α←␈↓	Would that be ok?
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓↓␈↓ βW{Several␈α∞simple␈α∞heuristics␈α∂are␈α∞invoked␈α∞to␈α∂select␈α∞the␈α∞next␈α∂most␈α∞likely
␈↓ α←␈↓↓␈↓ βWinterpretation,␈αbased␈αon␈αthe␈αEnglish␈αsource␈αof␈αthis␈αclause␈αand␈αthe␈αset
␈↓ α←␈↓␈↓5-2␈↓ εoDEBUGGING EXAMPLE CONTINUED    71␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βWof␈α
interpretations␈α∞already␈α
rejected.␈α∞ In␈α
this␈α
case,␈α∞the␈α
second␈α∞choice␈α
is
␈↓ α←␈↓↓␈↓ βWcorrect.}

␈↓ α←␈↓	Then you'll probably want to delete:
␈↓ α←␈↓	        2) The infection with the organism was acquired while
␈↓ α←␈↓	           the patient was hospitalized
␈↓ α←␈↓	Correct?
␈↓ α←␈↓	++** ␈↓αY␈↓	

␈↓ α←␈↓↓␈↓ βW{␈↓¬TEIRESIAS␈↓↓␈αenforces␈αconsistency␈α
constraints␈αon␈αits␈αinterpretation␈α
of␈αthe
␈↓ α←␈↓↓␈↓ βWEnglish␈α∩text.␈α∩ One␈α∪of␈α∩these␈α∩constraints␈α∪indicates␈α∩that␈α∩if␈α∪the␈α∩new
␈↓ α←␈↓↓␈↓ βWinterpretation␈αof␈α
clause␈α#1␈αis␈α
correct,␈αthen␈αclause␈α
#2␈αmust␈αhave␈α
arisen
␈↓ α←␈↓↓␈↓ βWspuriously.␈α∂ That␈α∂is,␈α∂the␈α∂system␈α∞realizes␈α∂it␈α∂cannot␈α∂get␈α∂␈↓&both␈↓)αβ␈α∂the␈α∞new
␈↓ α←␈↓↓␈↓ βWversion␈αof␈αclause␈α#1␈αand␈αthe␈αold␈αclause␈α#2␈αfrom␈αthe␈αsame␈αline␈αof␈αtext,
␈↓ α←␈↓↓␈↓ βWand indicates this to the expert.}

␈↓ α←␈↓	Anything else?
␈↓ α←␈↓	++**␈↓α D 3␈↓	
␈↓ α←␈↓	clause(s) 3 deleted.

␈↓ α←␈↓↓␈↓ βW{Realizing␈αthat␈αthere␈αwas␈αanother␈αspurious␈αclause,␈αthe␈αexpert␈αdeletes
␈↓ α←␈↓↓␈↓ βWclause␈α#3,␈αbut␈αthen␈αhas␈αsecond␈αthoughts␈αabout␈αthe␈αone␈αhe␈αchose,␈αso␈αhe
␈↓ α←␈↓↓␈↓ βWchecks, using the ``?'' command.}

␈↓ α←␈↓	Anything else?
␈↓ α←␈↓	++**␈↓α ?␈↓	
␈↓ α←␈↓	     1 - the infection is primary-bacteremia
␈↓ α←␈↓	 and 2 -   <deleted>
␈↓ α←␈↓	 and 3 -   <deleted>
␈↓ α←␈↓	 and 4 - the organism is not normally found at this site

␈↓ α←␈↓↓␈↓ βW{He␈α
then␈α
discovers␈α
that␈α
he␈α
has␈α
typed␈α
the␈α
wrong␈α
number,␈α∞undoes␈α
the
␈↓ α←␈↓↓␈↓ βWdamage, and deletes the one initially intended.}

␈↓ α←␈↓	Anything else?
␈↓ α←␈↓	++**␈↓α U 3␈↓	
␈↓ α←␈↓	clause(s) 3 undeleted.

␈↓ α←␈↓	Anything else?
␈↓ α←␈↓	++**␈↓α D 4␈↓	
␈↓ α←␈↓	clause(s) 4 deleted.

␈↓ α←␈↓↓␈↓ βW{There␈α∃is␈α∃one␈α∀last␈α∃correction␈α∃to␈α∃be␈α∀made␈α∃(all␈α∃commands␈α∃can␈α∀be
␈↓ α←␈↓↓␈↓ βWabbreviated to a single letter)... }

␈↓ α←␈↓	Anything else?
␈↓ α←␈↓	++**␈↓α C 3␈↓	
␈↓ α←␈↓	Suppose (3) were changed to read-
␈↓ α←␈↓	        3 - The site of the culture is one of: those sites that
␈↓ α←␈↓␈↓72    KNOWLEDGE ACQUISITION I␈↓ 
#5-2␈↓

␈↓"β␈↓ α←␈↓	            are normally sterile
␈↓ α←␈↓	Would that be ok?
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓	Anything else?
␈↓ α←␈↓	++**␈↓α ?␈↓	
␈↓ α←␈↓	     1 - the infection is primary-bacteremia
␈↓ α←␈↓	 and 2 -   <deleted>
␈↓ α←␈↓	 and 3 - the site of the culture is one of: those sites that
␈↓ α←␈↓	         are normally sterile
␈↓ α←␈↓	 and 4 -   <deleted>

␈↓ α←␈↓	++**␈↓α OK␈↓	

␈↓ α←␈↓↓␈↓ βW{... and at last he's happy with the rule.}

␈↓ α←␈↓	Change action?
␈↓ α←␈↓	++**␈↓α N␈↓	

␈↓ α←␈↓	This is my understanding of your rule:
␈↓ α←␈↓	RULE383
␈↓ α←␈↓	-------
␈↓ α←␈↓	  If:  1) The infection is primary-bacteremia, and
␈↓ α←␈↓	       2) The site of the culture is one of: those sites that
␈↓ α←␈↓	          are normally sterile
␈↓ α←␈↓	  Then:  There is strongly suggestive evidence (.8) that the
␈↓ α←␈↓	         category of the organism is enterobacteriaceae


␈↓ α←␈↓	Okay? (Yes or No)
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓↓␈↓ βW{This␈αis␈αthe␈αclassic␈αtest␈αof␈αcomprehension: ␈αThe␈αoriginal␈αEnglish␈αtext
␈↓ α←␈↓↓␈↓ βWhas␈αbeen␈αturned␈αinto␈αa␈αrule␈αcoded␈αin␈α␈↓¬LISP␈↓↓;␈αthe␈α␈↓¬LISP␈↓↓␈αis␈αthen␈αtranslated
␈↓ α←␈↓↓␈↓ βWback␈α⊗into␈α⊗English␈α⊗for␈α∃approval.␈α⊗ Compare␈α⊗the␈α⊗results␈α⊗with␈α∃the
␈↓ α←␈↓↓␈↓ βWexpert's␈α⊂original␈α⊂input.␈α⊂ Now␈α⊂that␈α⊂the␈α⊂expert␈α⊂is␈α⊂satisfied␈α⊂that␈α⊂the
␈↓ α←␈↓↓␈↓ βWsystem␈α⊂understands␈α⊂the␈α⊃rule␈α⊂as␈α⊂he␈α⊃has␈α⊂entered␈α⊂it,␈α⊃␈↓¬TEIRESIAS␈↓↓␈α⊂checks
␈↓ α←␈↓↓␈↓ βWseveral things about the rule to see if ␈↓&it␈↓)αβ is satisfied.}

␈↓ α←␈↓	I hate to criticize, Dr. Davis, but did you know that most
␈↓ α←␈↓	rules about what the category of a organism might be, that
␈↓ α←␈↓	mention-
␈↓ α←␈↓	            the site of a culture
␈↓ α←␈↓	            the infection
␈↓ α←␈↓	ALSO mention-
␈↓ α←␈↓	     [A] - the portal of entry of a organism
␈↓ α←␈↓	Shall I try to write a clause to account for [A] ?
␈↓ α←␈↓	++**␈↓α Y␈↓	
␈↓ α←␈↓␈↓5-2␈↓ εoDEBUGGING EXAMPLE CONTINUED    73␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βW{The␈αmechanism␈α
that␈αmakes␈αpossible␈α
this␈αsort␈α
of␈α``second␈αguessing''␈α
by
␈↓ α←␈↓↓␈↓ βW␈↓¬TEIRESIAS␈↓↓␈αis␈αexplained␈αin␈αdetail␈αbelow.␈αBriefly,␈αthe␈αsystem␈αmaintains␈αa
␈↓ α←␈↓↓␈↓ βWmodel␈α
of␈α
the␈α
knowledge␈αbase.␈α
 In␈α
checking␈α
to␈αsee␈α
how␈α
this␈α
new␈αpiece␈α
of
␈↓ α←␈↓↓␈↓ βWinformation␈α
``fits␈α
in''␈α
to␈α
that␈α
model,␈α
it␈α
discovers␈α
that␈α
its␈α
expectations
␈↓ α←␈↓↓␈↓ βWhave␈α⊗been␈α↔only␈α⊗partially␈α↔met.␈α⊗ It␈α⊗indicates␈α↔to␈α⊗the␈α↔expert␈α⊗what
␈↓ α←␈↓↓␈↓ βWexpectations␈α
were␈αleft␈α
unfulfilled␈α
and␈αasks␈α
if␈α
it␈αshould␈α
attempt␈αto␈α
fill
␈↓ α←␈↓↓␈↓ βWin the missing parts.}

␈↓ α←␈↓	How about-
␈↓ α←␈↓	     [A] The portal of entry of the organism is gi
␈↓ α←␈↓	Ok?
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈α∞expert␈α∂agreed␈α∞to␈α∂the␈α∞inclusion␈α∞of␈α∂a␈α∞clause␈α∂about␈α∞the␈α∂portal␈α∞of
␈↓ α←␈↓↓␈↓ βWentry,␈α∞this␈α∞determines␈α∞which␈α∞attribute␈α∞to␈α∞use.␈α∞ ␈↓¬TEIRESIAS␈↓↓␈α∞examines␈α∞its
␈↓ α←␈↓↓␈↓ βWmodel␈αof␈αthe␈α
knowledge␈αbase␈αto␈αsee␈α
which␈αpredicate␈αfunction␈α
is␈αlikely
␈↓ α←␈↓↓␈↓ βWto␈αbe␈αthe␈αright␈αone;␈α
the␈αmodel␈αevidently␈αsuggested␈αthe␈α
function␈α␈↓	SAME␈↓↓.
␈↓ α←␈↓↓␈↓ βWFinally,␈αin␈αorder␈αto␈αfix␈αthe␈αoriginal␈αbug,␈αit␈αmust␈αbe␈αpossible␈αto␈αinvoke
␈↓ α←␈↓↓␈↓ βWthe␈αnew␈αrule␈αfor␈α
this␈αpatient.␈α Since,␈αin␈α
answer␈αto␈αquestion␈α18␈α
of␈αthe
␈↓ α←␈↓↓␈↓ βWoriginal␈α
consultation␈α
(see␈α
Section␈α
3-10),␈α
the␈α
expert␈α
responded␈αthat␈α
the
␈↓ α←␈↓↓␈↓ βWportal␈αof␈αentry␈α
was␈α␈↓αGI␈↓↓,␈αthis␈α
is␈αa␈αplausible␈αguess␈α
at␈αthe␈αvalue␈α
to␈αuse.
␈↓ α←␈↓↓␈↓ βWThis␈α⊂demonstrates␈α∂another␈α⊂benefit␈α∂of␈α⊂the␈α∂``knowledge␈α⊂acquisition␈α∂in
␈↓ α←␈↓↓␈↓ βWcontext'' methodology described in chapter 4.}

␈↓ α←␈↓	Also, most rules about what the category of a organism might be
␈↓ α←␈↓	ALSO conclude about-
␈↓ α←␈↓	     [A] - the identity of a organism
␈↓ α←␈↓	Shall I try to write a clause to account for [A] ?
␈↓ α←␈↓	++**␈↓α N␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈α⊂system␈α⊂noticed␈α⊂another␈α⊂unfulfilled␈α⊂expectation␈α⊃concerning␈α⊂the
␈↓ α←␈↓↓␈↓ βWaction␈α∩part␈α∩of␈α∩the␈α∩rule,␈α∪but␈α∩the␈α∩expert␈α∩always␈α∩has␈α∩the␈α∪option␈α∩of
␈↓ α←␈↓↓␈↓ βWoverriding the system's suggestions, and does so.}

␈↓ α←␈↓	The rule now reads:
␈↓ α←␈↓	RULE383
␈↓ α←␈↓	-------
␈↓ α←␈↓	  If:  1) The infection is primary-bacteremia, and
␈↓ α←␈↓	       2) The site of the culture is one of: those sites that
␈↓ α←␈↓	          are normally sterile, and
␈↓ α←␈↓	       3) The portal of entry of the organism is gi
␈↓ α←␈↓	  Then:  There is strongly suggestive evidence (.8) that the
␈↓ α←␈↓	         category of the organism is enterobacteriaceae


␈↓ α←␈↓↓␈↓ βW{Now␈α
that␈α∞both␈α
the␈α∞expert␈α
and␈α∞␈↓¬TEIRESIAS␈↓↓␈α
are␈α∞satisfied␈α
with␈α∞the␈α
rule,
␈↓ α←␈↓↓␈↓ βWthere␈α
is␈α
a␈α
bit␈αof␈α
bookkeeping␈α
to␈α
do.␈α
First,␈αthe␈α
new␈α
rule␈α
is␈α
checked␈αby
␈↓ α←␈↓␈↓74    KNOWLEDGE ACQUISITION I␈↓ 
#5-2␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βWthe␈α∀system␈α∀(behind␈α∀the␈α∀scenes)␈α∀to␈α∀be␈α∀sure␈α∀that␈α∀there␈α∃is␈α∀nothing
␈↓ α←␈↓↓␈↓ βWobviously␈α∂wrong␈α∂with␈α⊂it--for␈α∂instance,␈α∂one␈α⊂of␈α∂the␈α∂tests␈α⊂insures␈α∂that
␈↓ α←␈↓↓␈↓ βWnone␈α
of␈αthe␈α
premise␈α
clauses␈αis␈α
currently␈α
known␈αto␈α
be␈α
false.␈αThis␈α
might
␈↓ α←␈↓↓␈↓ βWhave␈α⊃happened␈α∩if␈α⊃the␈α∩expert␈α⊃had␈α⊃mistakenly␈α∩typed␈α⊃␈↓	the␈α∩site␈α⊃is
␈↓ α←␈↓	␈↓ βWnot␈α∪one␈α∀of␈α∪the␈α∀sterile␈α∪sites␈↓↓.␈α∀ Had␈α∪this␈α∀been␈α∪the␈α∀case,␈α∪the
␈↓ α←␈↓↓␈↓ βWsystem␈α↔would␈α⊗have␈α↔noticed␈α↔at␈α⊗this␈α↔point␈α⊗that␈α↔that␈α↔clause␈α⊗was
␈↓ α←␈↓↓␈↓ βWguaranteed␈αto␈αfail␈αfor␈αthe␈αcurrent␈αpatient,␈αand␈αthe␈αexpert␈αwould␈αhave
␈↓ α←␈↓↓␈↓ βWbeen␈α∪given␈α∪the␈α∪option␈α∪of␈α∪returning␈α∪to␈α∪the␈α∪rule␈α∪editor␈α∪to␈α∪fix␈α∪the
␈↓ α←␈↓↓␈↓ βWproblem.␈α⊃ There␈α⊂are␈α⊃several␈α⊃other␈α⊂things␈α⊃the␈α⊂system␈α⊃checks␈α⊃at␈α⊂this
␈↓ α←␈↓↓␈↓ βWpoint,␈α∀but␈α∀this␈α∀rule␈α∀passes␈α∀all␈α∪the␈α∀tests,␈α∀and␈α∀hence␈α∀there␈α∀is␈α∪no
␈↓ α←␈↓↓␈↓ βWindication of any problem.}

␈↓ α←␈↓	Please describe briefly why you added (or changed) this rule.
␈↓ α←␈↓	Type an empty line when done.
␈↓ α←␈↓	++**␈↓αTHE SYSTEM FAILED TO REALIZE IT COULD CONCLUDE␈↓	
␈↓ α←␈↓	++**␈↓αCATEGORY, AND THIS ALLOWED RULE184 TO INCORRECTLY␈↓	
␈↓ α←␈↓	++**␈↓αCONCLUDE IDENTITY␈↓	
␈↓ α←␈↓	++**

␈↓ α←␈↓↓␈↓ βW{Over␈α
the␈αyears␈α
of␈α
␈↓¬MYCIN␈↓↓␈αdevelopment,␈α
the␈αexperts␈α
associated␈α
with␈αthe
␈↓ α←␈↓↓␈↓ βWproject␈α∪have␈α∪occasionally␈α∩looked␈α∪in␈α∪astonishment␈α∩at␈α∪a␈α∪rule␈α∩which,
␈↓ α←␈↓↓␈↓ βWthough␈α⊃it␈α⊂had␈α⊃been␈α⊂around␈α⊃for␈α⊂some␈α⊃time,␈α⊂seemed,␈α⊃superficially␈α⊂at
␈↓ α←␈↓↓␈↓ βWleast,␈α∞to␈α∞make␈α∞very␈α∞little␈α∂sense.␈α∞ The␈α∞question␈α∞then␈α∞arises␈α∞as␈α∂to␈α∞why
␈↓ α←␈↓↓␈↓ βWthat␈α
rule␈α
was␈α
added␈α
to␈α
the␈α
knowledge␈α
base.␈α
 To␈α
avoid␈α
this␈αproblem,␈α
the
␈↓ α←␈↓↓␈↓ βWexpert is required to explain why he wrote (or changed) the rule.
␈↓ α←␈↓↓␈↓ βWThere␈α∂are␈α∂also␈α∂a␈α∂number␈α∂of␈α∂other␈α∂bookkeeping-type␈α∂chores␈α⊂that␈α∂are
␈↓ α←␈↓↓␈↓ βWbeing␈αattended␈α
to␈αhere␈α
(again,␈αbehind␈α
the␈αscenes).␈α
 ␈↓¬TEIRESIAS␈↓↓␈αtakes␈α
care
␈↓ α←␈↓↓␈↓ βWof␈αtagging␈αthe␈αrule␈αwith␈αinformation␈αthat␈αwill␈αsimplify␈αkeeping␈α
track
␈↓ α←␈↓↓␈↓ βWof␈αthe␈α
large,␈αand␈αconstantly␈α
changing,␈αbody␈α
of␈αrules.␈α We␈α
will␈αsee␈αin␈α
a
␈↓ α←␈↓↓␈↓ βWmoment how that information is accessed.}

␈↓ α←␈↓	RULE383 has now been added to the knowledge base.

␈↓ α←␈↓	 I will now re-run the consultation to check the effectiveness
␈↓ α←␈↓	of your new rule. Get comfortable, since this may take a bit.
␈↓ α←␈↓	     Also, new questions may be asked as a result of this new
␈↓ α←␈↓	rule. If a question is printed, please answer it with reference
␈↓ α←␈↓	to this current patient.

␈↓ α←␈↓↓␈↓ βW{␈↓¬TEIRESIAS␈↓↓␈α∂now␈α∂invokes␈α∂the␈α∞performance␈α∂program␈α∂as␈α∂a␈α∂subprocess␈α∞to
␈↓ α←␈↓↓␈↓ βWmake␈α⊂sure␈α∂the␈α⊂bug␈α⊂is␈α∂fixed,␈α⊂and␈α∂to␈α⊂test␈α⊂for␈α∂any␈α⊂other␈α⊂side␈α∂effects.
␈↓ α←␈↓↓␈↓ βWSince␈α∂␈↓¬MYCIN␈↓↓␈α∂has␈α∞saved␈α∂all␈α∂the␈α∞answers␈α∂from␈α∂the␈α∂initial␈α∞consultation,
␈↓ α←␈↓↓␈↓ βWthe␈α∂process␈α∂is␈α∂fairly␈α∂inexpensive.␈α∂ The␈α∂expert␈α∂is␈α∂warned␈α⊂about␈α∂new
␈↓ α←␈↓↓␈↓ βWquestions␈αbecause␈αhis␈αnew␈αrule␈αmay␈αmention␈αan␈αattribute␈αthat␈αwas␈αnot
␈↓ α←␈↓↓␈↓ βWpreviously traced, or it may indirectly cause the tracing of one.
␈↓ α←␈↓↓␈↓ βWWhile␈α⊂there␈α⊂are␈α⊃some␈α⊂cases␈α⊂in␈α⊂which␈α⊃it␈α⊂is␈α⊂possible␈α⊂to␈α⊃compute␈α⊂the
␈↓ α←␈↓␈↓5-2␈↓ εoDEBUGGING EXAMPLE CONTINUED    75␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βWincremental␈α⊂effect␈α⊂on␈α⊂a␈α⊂given␈α∂consultation␈α⊂of␈α⊂adding␈α⊂a␈α⊂single␈α∂new
␈↓ α←␈↓↓␈↓ βWrule␈α
to␈α
the␈αknowledge␈α
base,␈α
these␈αare␈α
rare␈α
enough␈αthat␈α
in␈α
the␈αlong␈α
run
␈↓ α←␈↓↓␈↓ βWit is computationally cheaper to simply rerun the consultation.}

␈↓ α←␈↓	18-DEC-75 16:30:46

␈↓ α←␈↓	   --------PATIENT-281--------

␈↓ α←␈↓	      --------INFECTION-1--------

␈↓ α←␈↓	         --------CULTURE-1--------

␈↓ α←␈↓	            --------ORGANISM-1--------

␈↓ α←␈↓↓␈↓ βW{The␈α⊗answers␈α↔to␈α⊗all␈α⊗questions␈α↔were␈α⊗available␈α⊗internally,␈α↔so␈α⊗the
␈↓ α←␈↓↓␈↓ βWperformance␈α
program␈α
asks␈α
nothing␈α
new.␈α
(The␈α
headings␈α
are␈αprinted␈α
to
␈↓ α←␈↓↓␈↓ βWlet the expert follow the progress of the consultation.)}

␈↓ α←␈↓	My therapy recommendation will be based on the following
␈↓ α←␈↓	probable infection(s) and potential causative organism(s):

␈↓ α←␈↓	INFECTION-1 is PRIMARY-BACTEREMIA
␈↓ α←␈↓	    <Item 1>  E.COLI [ORGANISM-1]
␈↓ α←␈↓	    <Item 2>  KLEBSIELLA-PNEUMONIAE [ORGANISM-1]
␈↓ α←␈↓	    <Item 3>  PROTEUS-NON-MIRABILIS [ORGANISM-1]


␈↓ α←␈↓	Well, it appears that things are straightened out.

␈↓ α←␈↓↓␈↓ βW{The␈αnew␈αconclusions␈αare␈αthe␈αones␈αthe␈αexpert␈αindicated␈αearlier␈αas␈αthe
␈↓ α←␈↓↓␈↓ βWcorrect␈α∂answers.␈α∂ This␈α∂does␈α∂not␈α∂go␈α∂unnoticed␈α∂by␈α∂␈↓¬TEIRESIAS␈↓↓,␈α∂which,␈α∂as
␈↓ α←␈↓↓␈↓ βWbefore,␈α
interrupts␈α
the␈α∞consultation␈α
at␈α
this␈α∞point␈α
to␈α
check␈α∞the␈α
validity
␈↓ α←␈↓↓␈↓ βWof the answers produced.}

␈↓ α←␈↓	Are all these values correct now?
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓↓␈↓ βW{As␈α⊂always,␈α∂␈↓¬TEIRESIAS␈↓↓␈α⊂gives␈α⊂the␈α∂expert␈α⊂the␈α⊂last␈α∂word␈α⊂and␈α⊂waits␈α∂for
␈↓ α←␈↓↓␈↓ βWapproval before proceeding.}

␈↓ α←␈↓	Good, then the problem's been fixed.
␈↓ α←␈↓	     ...continuing with therapy selection now.

␈↓ α←␈↓↓␈↓ βW{At␈αthis␈αpoint␈α
␈↓¬TEIRESIAS␈↓↓␈αallows␈αthe␈α
performance␈αprogram␈αsubprocess␈α
to
␈↓ α←␈↓↓␈↓ βWcontinue␈α∩with␈α∩the␈α∪normal␈α∩therapy-selection␈α∩phase.␈α∩ This␈α∪has␈α∩been
␈↓ α←␈↓↓␈↓ βWomitted here for brevity.
␈↓ α←␈↓↓␈↓ βWThe␈αreader␈α
may␈αbe␈α
wondering␈αwhy␈αthree␈α
errors␈αdisappeared␈α
with␈αthe
␈↓ α←␈↓↓␈↓ βWaddition of one rule. This is explained schematically below:
␈↓"β␈↓ α←␈↓␈↓76    KNOWLEDGE ACQUISITION I␈↓ 
#5-2␈↓

␈↓"β␈↓ α←␈↓¬RULE116 **
␈↓ α←␈↓¬RULE050 **
␈↓ α←␈↓¬RULE383  --->  CATEGORY  --- ** RULE184 ** --->  IDENTITY =  PSEUDOMONAS-AERUGINOSA

␈↓ α←␈↓¬                         ------ RULE084 ------>  IDENTITY =  KLEBSIELLA-PNEUMONIAE
␈↓ α←␈↓¬                                                             PROTEUS-NON-MIRABILIS
␈↓ α←␈↓↓␈↓ βWAs␈α
before,␈α
116␈α
and␈α
50␈α
have␈α
failed␈α
(hence␈α
the␈α
**s),␈α
but␈α
now␈α∞the␈α
new
␈↓ α←␈↓↓␈↓ βWrule,␈α∂383,␈α∂has␈α∂succeeded,␈α∂supplying␈α∂a␈α∂value␈α∂for␈α∂category.␈α∂ In␈α∂turn,
␈↓ α←␈↓↓␈↓ βW184␈α∪now␈α∀fails,␈α∪as␈α∀hoped,␈α∪and␈α∀thus␈α∪no␈α∀conclusion␈α∪is␈α∀made␈α∪about
␈↓ α←␈↓↓␈↓ βWpseudomonas-aeruginosa.
␈↓ α←␈↓↓␈↓ βWHowever,␈α
rule␈α
84,␈αwhich␈α
requires␈α
a␈α
value␈αof␈α
category␈α
and␈α
failed␈αlast
␈↓ α←␈↓↓␈↓ βWtime␈α
because␈α
the␈α
system␈α
was␈α∞unable␈α
to␈α
deduce␈α
one,␈α
can␈α
now␈α∞fire␈α
and
␈↓ α←␈↓↓␈↓ βWyields the desired, missing identities.}

␈↓ α←␈↓	␈↓ ¬7********************


␈↓ α←␈↓↓␈↓ βW{All␈α~of␈α~the␈α→bookkeeping␈α~that␈α~was␈α→done␈α~is␈α~accessible␈α~via␈α→the
␈↓ α←␈↓↓␈↓ βW␈↓INFO␈↓↓rmation␈α∂command␈α∂added␈α∂to␈α∂the␈α∂question-answering␈α∂facility␈α∞of
␈↓ α←␈↓↓␈↓ βWthe␈α
performance␈α
program.␈α
 This␈α
gives␈α
the␈α
expert␈α
the␈α
background␈αfor
␈↓ α←␈↓↓␈↓ βWany rule in the system.}

␈↓ α←␈↓	** ␈↓αINFO 383␈↓	

␈↓ α←␈↓	RULE383
␈↓ α←␈↓	-------
␈↓ α←␈↓	  was written by Dr. Davis
␈↓ α←␈↓	  on December 18, 1975
␈↓ α←␈↓	  for the following reason:
␈↓ α←␈↓	            THE SYSTEM FAILED TO REALIZE IT COULD CONCLUDE
␈↓ α←␈↓	            CATEGORY, AND THIS ALLOWED RULE184 TO INCORRECTLY
␈↓ α←␈↓	            CONCLUDE IDENTITY.

␈↓ α←␈↓	  for patient 281
␈↓ α←␈↓	       [who was described as follows:
␈↓ α←␈↓	            CLASSIC CASE OF GRAM NEGATIVE ROD INFECTION FOR A
␈↓ α←␈↓	            PATIENT WITH A NON-NOSOCOMIAL DISEASE]
␈↓ α←␈↓␈↓5-3␈↓ πbRULE MODEL OVERVIEW    77␈↓

␈↓"β␈↓ α←␈↓␈↓α5-3    RULE MODEL OVERVIEW␈↓
␈↓ α←␈↓␈↓ β?The␈α
utility␈α
of␈α
knowledge␈α
acquisition␈α
in␈α
context␈α
should␈α
now␈α
be␈αclear.
␈↓ α←␈↓In␈α⊗this␈α∃example␈α⊗we␈α⊗have␈α∃seen␈α⊗how␈α⊗the␈α∃availability␈α⊗of␈α⊗the␈α∃contextual
␈↓ α←␈↓information␈αhas␈αmade␈αit␈αmuch␈αeasier␈αfor␈αthe␈αexpert␈αto␈αspecify␈αthe␈αknowledge
␈↓ α←␈↓required␈αand␈αhow␈αit␈αhas␈αsupported,␈αas␈αwell,␈αmany␈αof␈αthe␈α``intelligent''␈αactions
␈↓ α←␈↓taken by the program.
␈↓ α←␈↓␈↓ β?One␈α∩of␈α∩the␈α∩uses␈α⊃of␈α∩this␈α∩contextual␈α∩information␈α⊃is␈α∩as␈α∩a␈α∩source␈α⊃of
␈↓ α←␈↓expectations␈α
about␈α
the␈α
content␈α
of␈α∞the␈α
expert's␈α
new␈α
rule.␈α
 The␈α
concept␈α∞of␈α
an
␈↓ α←␈↓``expectation''␈α
is␈α
precisely␈α
defined␈α
in␈α
␈↓¬TEIRESIAS␈↓,␈α
and␈α
in␈α
order␈α
to␈α
provide␈α
some
␈↓ α←␈↓perspective␈α∀on␈α∀its␈α∀intellectual␈α∀origins␈α∀and␈α∀development,␈α∀we␈α∀consider␈α∀the
␈↓ α←␈↓concept of ␈↓↓models␈↓␈↓
1␈↓ and ␈↓↓model-based understanding␈↓.
␈↓ α←␈↓␈↓ β?The␈α
use␈αof␈α
models␈αin␈α
computer␈αprograms␈α
as␈αa␈α
guide␈αto␈α
understanding
␈↓ α←␈↓dates␈αback␈αto␈αearly␈αefforts␈αin␈αAI.␈αThe␈αauthor's␈αfirst␈αencounter␈αwith␈αit␈αwas␈αin
␈↓ α←␈↓work␈α~on␈α→computer␈α~vision,␈α→and␈α~since␈α→this␈α~offers␈α→a␈α~particularly␈α→clear
␈↓ α←␈↓introduction␈α⊂to␈α⊃the␈α⊂ideas␈α⊂involved,␈α⊃we␈α⊂digress␈α⊂for␈α⊃a␈α⊂moment␈α⊃to␈α⊂introduce
␈↓ α←␈↓some␈α∩useful␈α∩concepts␈α∩from␈α∩that␈α∩field.␈α∩ (The␈α∩treatment␈α∩here␈α∩is␈α∩necessarily
␈↓ α←␈↓superficial.␈α⊂For␈α⊂a␈α⊂broader␈α⊂review␈α∂of␈α⊂current␈α⊂work,␈α⊂see␈α⊂[Barrow75];␈α⊂for␈α∂an
␈↓ α←␈↓interesting␈α∂introduction␈α∂to␈α∂the␈α∂capabilities␈α∂of␈α∂the␈α∂human␈α∂visual␈α∂system,␈α∞see
␈↓ α←␈↓[Gregory66].)

␈↓ α←␈↓␈↓α5-3-1    Perspective:  Model-based computer vision␈↓
␈↓ α←␈↓␈↓ β?Computer␈α
interpretation␈α
of␈α
a␈αscene␈α
typically␈α
starts␈α
with␈αdigitization,␈α
in
␈↓ α←␈↓which␈α
the␈αpicture␈α
is␈α
turned␈αinto␈α
a␈αlarge␈α
array␈α
of␈αnumbers␈α
indicating␈αthe␈α
light
␈↓ α←␈↓intensity␈αat␈αeach␈αpoint.␈α The␈αtask␈αis␈αthen␈αto␈αdetermine␈αwhich␈αphysical␈αobjects
␈↓ α←␈↓in␈αthe␈αscene␈αproduced␈αeach␈αof␈αthe␈αdifferent␈αregions␈αof␈αintensity␈αvalues␈αin␈αthe
␈↓ α←␈↓array.␈α Among␈αthe␈αtools␈αused␈αare␈αvarious␈αtypes␈αof␈αedge␈αdetectors.␈α These␈αare
␈↓ α←␈↓mathematical␈α∞operators␈α∞that␈α∞are␈α∞sensitive␈α∂to␈α∞certain␈α∞sorts␈α∞of␈α∞changes␈α∂in␈α∞the
␈↓ α←␈↓intensity␈α
level␈α
values,␈αusually␈α
sharp␈α
gradients␈αor␈α
step␈α
functions,␈αwhich␈α
suggest
␈↓ α←␈↓the end of one region and the beginning of another.
␈↓ α←␈↓␈↓ β?One␈α⊃of␈α⊃the␈α⊃first␈α⊂significant␈α⊃stumbling␈α⊃blocks␈α⊃encountered␈α⊃in␈α⊂scene
␈↓ α←␈↓understanding␈α∂was␈α∂the␈α∂presence␈α∂of␈α∂shadows␈α∂and␈α∂noise.␈↓
2␈↓␈α∂Real␈α⊂scenes␈α∂rarely
␈↓ α←␈↓contain␈αall␈αthe␈αdetail␈αseen␈α(or␈αinterpreted)␈αby␈αthe␈αhuman␈αeye/brain␈αprocessor.
␈↓ α←␈↓Many␈α∞of␈α∞the␈α∞edges␈α∞of␈α∞objects,␈α∞for␈α∞instance,␈α∞are␈α∞lost␈α∞in␈α∞shadows;␈α∞these␈α
same
␈↓ α←␈↓shadows␈α↔often␈α↔suggest␈α↔false␈α↔edges␈α↔as␈α↔well.␈α↔ As␈α↔a␈α↔result,␈α_the␈α↔uniform
␈↓ α←␈↓application␈α∞of␈α∞an␈α∞edge␈α∞detector␈α∞results␈α∞in␈α∞a␈α∞large␈α∞collection␈α∞of␈α∂``edges,''␈α∞only

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[1]␈α∀The␈α∀concept␈α∀of␈α∪a␈α∀model␈α∀has␈α∀been␈α∪an␈α∀integral␈α∀part␈α∀of␈α∀science␈α∪and
␈↓ α←␈↓philosophy␈α⊂for␈α⊂quite␈α⊂some␈α⊂time.␈α⊂The␈α∂work␈α⊂described␈α⊂here␈α⊂can␈α⊂be␈α⊂seen␈α∂in
␈↓ α←␈↓terms␈α
of␈α∞the␈α
standard␈α∞definitions␈α
that␈α
have␈α∞evolved,␈α
and␈α∞it␈α
fits␈α∞quite␈α
easily
␈↓ α←␈↓into␈α∞the␈α∞frameworks␈α
that␈α∞have␈α∞been␈α∞developed.␈α
 Since␈α∞the␈α∞literature␈α∞on␈α
the
␈↓ α←␈↓subject␈αis␈α
extensive,␈αwe␈αwill␈α
not␈αdeal␈α
with␈αit␈αhere.␈α
For␈αa␈α
useful␈αintroduction,
␈↓ α←␈↓see [Harre70], especially chapter 2.

␈↓ α←␈↓[2]␈α
Noise␈α
can␈α
be␈α
electronic,␈α
can␈α
be␈α
due␈α
to␈α
digitization␈α
effects,␈α
or␈α
can␈α
simply␈α
be
␈↓ α←␈↓the result of smudges or dirt on the objects in the scene.
␈↓ α←␈↓␈↓78    KNOWLEDGE ACQUISITION I␈↓ 
#5-3␈↓

␈↓"β␈↓ α←␈↓some␈α∞of␈α∞which␈α∞are␈α∞real.␈α∞ It␈α∞is␈α∞in␈α∞part␈α∞because␈α∞human␈α∞vision␈α∞is␈α∞an␈α∂eye␈α∞␈↓↓and␈↓
␈↓ α←␈↓brain␈α⊃process␈α⊃that␈α⊃we␈α⊃are␈α⊃not␈α⊃similarly␈α⊃confused--we␈α⊃are␈α⊃able␈α∩to␈α⊃classify
␈↓ α←␈↓objects␈αbeing␈αviewed␈αand,␈αbased␈αon␈αthis␈αclassification,␈αcan␈αignore␈αfalse␈αedges
␈↓ α←␈↓and supply missing ones.
␈↓ α←␈↓␈↓ β?Some␈αof␈αthe␈αfirst␈αvision␈αwork␈αto␈αtake␈αthese␈αproblems␈αinto␈αaccount␈α
was
␈↓ α←␈↓done␈α⊂by␈α⊂Falk␈α⊂[Falk70],␈α⊂building␈α⊂on␈α⊂the␈α⊂work␈α⊂of␈α⊂Roberts␈α⊃[Roberts63]␈α⊂and
␈↓ α←␈↓Guzman␈α∂[Guzman68].␈α⊂ The␈α∂key␈α⊂elements␈α∂of␈α⊂this␈α∂work␈α⊂that␈α∂are␈α⊂of␈α∂interest
␈↓ α←␈↓here␈αare␈α
(␈↓↓i␈↓)  a␈αset␈αof␈α
␈↓↓prototypes␈↓␈α(or␈αmodels)␈α
that␈α␈↓↓embody␈αknowledge␈↓␈α
about␈αthe
␈↓ α←␈↓objects␈α
being␈α
viewed␈αand␈α
(␈↓↓ii␈↓)  the␈α
use␈αof␈α
these␈α
prototypes␈αas␈α
a␈α
way␈αof␈α
␈↓↓guiding
␈↓ α←␈↓↓the process of understanding␈↓.
␈↓ α←␈↓␈↓ β?Falk's␈α
system␈α
(like␈α
many␈α
others)␈α
viewed␈α
scenes␈α
containing␈α
a␈αnumber␈α
of
␈↓ α←␈↓children's␈αblocks␈α
selected␈αfrom␈α
a␈αknown␈α
set␈αof␈α
possible␈αobjects.␈αHis␈α
prototypes
␈↓ α←␈↓were␈α
similar␈αto␈α
wire␈αframe␈α
models␈αof␈α
each␈α
of␈αthe␈α
blocks␈αand␈α
thus␈αembodied␈α
a
␈↓ α←␈↓``high-level''␈α⊃description␈α⊃that␈α⊃indicated␈α⊃the␈α⊃structure␈α⊃of␈α⊃each␈α⊃object.␈α⊃ This
␈↓ α←␈↓compact,␈α∀high-level␈α∀description␈α∀was␈α∀then␈α∀used␈α∀to␈α∀guide␈α∀the␈α∀lower␈α∪level
␈↓ α←␈↓processing␈αof␈αthe␈αintensity␈αvalues.␈α
 After␈αdetecting␈αa␈αnumber␈αof␈αedges␈α
in␈αthe
␈↓ α←␈↓scene,␈α
his␈α
system␈αattempted␈α
to␈α
fit␈αa␈α
model␈α
to␈α
each␈αcollection␈α
of␈α
edges␈αand␈α
then
␈↓ α←␈↓used␈αthis␈αmodel␈αas␈αa␈αguide␈α
to␈αfurther␈αprocessing.␈α If,␈αfor␈αinstance,␈α
the␈αmodel
␈↓ α←␈↓accounted␈αfor␈αall␈αbut␈αone␈αof␈αthe␈αlines␈αin␈αa␈αregion,␈αthis␈αwould␈αsuggest␈αthat␈αthe
␈↓ α←␈↓extra␈α⊃line␈α⊃might␈α⊃be␈α⊂spurious.␈α⊃ If␈α⊃the␈α⊃model␈α⊂fit␈α⊃well␈α⊃except␈α⊃for␈α⊃some␈α⊂line
␈↓ α←␈↓missing␈αfrom␈αthe␈α
scene,␈αthis␈αwas␈αa␈α
good␈αhint␈αthat␈α
a␈αline␈αhad␈αbeen␈α
overlooked
␈↓ α←␈↓and also an indication of where to go looking for it.
␈↓ α←␈↓␈↓ β?One␈αother␈αidea␈α
necessary␈αto␈αcomplete␈αa␈α
general␈αperspective␈αon␈αthe␈α
use
␈↓ α←␈↓of␈α⊂models␈α⊂to␈α⊂guide␈α⊂understanding␈α⊂is␈α⊂described␈α⊂quite␈α⊂well␈α⊂by␈α⊃a␈α⊂distinction
␈↓ α←␈↓drawn␈α∩in␈α∩[Baumgart74].␈α⊃ As␈α∩he␈α∩points␈α⊃out␈α∩(also␈α∩in␈α⊃the␈α∩context␈α∩of␈α⊃scene
␈↓ α←␈↓understanding,␈α∂but␈α∂it␈α∂generalizes␈α∂easily),␈α∂three␈α∂different␈α∂approaches␈α∂can␈α∂be
␈↓ α←␈↓identified:

␈↓ α←␈↓␈↓ ββ(1)␈↓ β?␈↓↓Verification␈↓.␈α" ``Top␈α"down''␈α"or␈α#completely␈α"model-driven
␈↓ α←␈↓␈↓ β?processing␈α∪concentrates␈α∪on␈α∪verifying␈α∪the␈α∪details␈α∪in␈α∪a␈α∩scene
␈↓ α←␈↓␈↓ β?about which much is already known.

␈↓ α←␈↓␈↓ ββ(2)␈↓ β?␈↓↓Description␈↓.␈α∞ ``Bottom␈α∞up''␈α∞or␈α∞data-driven␈α∞processing␈α∞involves
␈↓ α←␈↓␈↓ β?searching␈αan␈αunfamiliar␈αscene␈αfor␈αany␈αsuggestive␈αfeatures␈αthat
␈↓ α←␈↓␈↓ β?might indicate what is present.

␈↓ α←␈↓␈↓ ββ(3)␈↓ β?␈↓↓Recognition␈↓.␈α
 ``Bottom␈α
up␈α∞into␈α
a␈α
prejudiced␈α
top''␈α∞is␈α
processing
␈↓ α←␈↓␈↓ β?where␈α⊃there␈α⊃is␈α⊃available␈α⊃some␈α⊃idea␈α⊃of␈α⊃what␈α⊃to␈α∩expect,␈α⊃and
␈↓ α←␈↓␈↓ β?where␈αperhaps␈αa␈αsmall␈αset␈αof␈αpossibilities␈αcan␈αbe␈αspecified,␈αbut
␈↓ α←␈↓␈↓ β?none␈α
definitely.␈α The␈α
data␈α
is␈αallowed␈α
to␈αsuggest␈α
interpretations,
␈↓ α←␈↓␈↓ β?but␈α∂the␈α∞set␈α∂of␈α∂items␈α∞to␈α∂be␈α∂considered␈α∞is␈α∂constrained␈α∂to␈α∞those
␈↓ α←␈↓␈↓ β?expected.

␈↓ α←␈↓The␈α
first␈α
approach␈αis␈α
useful␈α
when␈αthe␈α
contents␈α
of␈αthe␈α
scene␈α
are␈α
known␈αand
␈↓ α←␈↓when␈αthe␈αtask␈αconcerns␈αonly␈αsettling␈αdetails␈αof␈αorientation␈αor␈αprecise␈αlocation.
␈↓ α←␈↓␈↓5-3␈↓ πbRULE MODEL OVERVIEW    79␈↓

␈↓"β␈↓ α←␈↓The␈αsecond␈αcan␈α
be␈αused␈αas␈α
the␈αinitial␈αapproach␈α
to␈αan␈αunfamiliar␈α
scene,␈αand
␈↓ α←␈↓the␈α∂third␈α∂is␈α∂relevant␈α∞if␈α∂there␈α∂is␈α∂some␈α∂idea␈α∞of␈α∂what␈α∂is␈α∂present,␈α∂but␈α∞nothing
␈↓ α←␈↓definite.␈α∩ (Of␈α⊃course,␈α∩all␈α⊃three␈α∩might␈α⊃well␈α∩be␈α⊃used␈α∩on␈α⊃the␈α∩same␈α∩scene␈α⊃at
␈↓ α←␈↓different points in the processing.)
␈↓ α←␈↓␈↓ β?Finally,␈αnote␈α
that␈αone␈α
of␈αthe␈α
long␈αrecognized␈α
potential␈αweaknesses␈αof␈α
a
␈↓ α←␈↓model-based␈αapproach␈αis␈αthe␈α
dependence␈αon␈αa␈αfixed␈α
set␈αof␈αmodels.␈α Since␈α
the
␈↓ α←␈↓scope␈α
of␈αthe␈α
program's␈α
``understanding''␈αof␈α
the␈α
world␈αmay␈α
be␈α
constrained␈αby
␈↓ α←␈↓the␈α
number␈α
of␈α
models␈α
it␈α
has,␈α
``...it␈αwould␈α
be␈α
desirable␈α
[for␈α
a␈α
program]␈α
to␈αbe
␈↓ α←␈↓able to `learn' new structural descriptions of models'' [Falk70].

␈↓ α←␈↓␈↓α5-3-2    Rule models:  Overview␈↓
␈↓ α←␈↓␈↓ β?For␈α↔rule␈α↔acquisition␈α_the␈α↔primary␈α↔problem␈α_to␈α↔be␈α↔solved␈α_is␈α↔the
␈↓ α←␈↓interpretation␈α⊃of␈α⊃the␈α⊃new␈α⊃rule.␈α⊃Note␈α⊃that␈α⊃this␈α⊃problem␈α⊃can␈α⊃be␈α⊃viewed␈α⊃as
␈↓ α←␈↓analogous␈αto␈αa␈αproblem␈αin␈αperception: ␈α(a)  There␈αis␈αa␈αsignal␈αto␈αbe␈αprocessed
␈↓ α←␈↓(text);␈α(b)␈αthe␈αsignal␈αis␈αnoisy␈α(``noise''␈αwords␈αthat␈αconvey␈αno␈αinformation);␈αand
␈↓ α←␈↓(c)␈α⊂there␈α⊃is␈α⊂a␈α⊃large␈α⊂amount␈α⊂of␈α⊃context␈α⊂available␈α⊃(from␈α⊂tracking␈α⊃down␈α⊂the
␈↓ α←␈↓error)␈α
on␈α
which␈α
to␈α
base␈α
expectations␈α
about␈α
the␈α
signal␈α
content.␈α
 It␈α
would␈α
prove
␈↓ α←␈↓useful,␈α
then,␈αto␈α
have␈αsomething␈α
analogous␈αto␈α
Falk's␈αprototypes,␈α
something␈αto
␈↓ α←␈↓capture␈α
the␈α
``structure''␈α
of␈α
the␈αphysician's␈α
reasoning,␈α
that␈α
would␈α
offer␈α
a␈αway␈α
of
␈↓ α←␈↓expressing␈α
expectations␈α
and␈α
guiding␈αthe␈α
interpretation␈α
of␈α
his␈α
statement␈αof␈α
the
␈↓ α←␈↓rule.␈α∂ In␈α∂short,␈α∂we␈α∂are␈α⊂asking␈α∂for␈α∂the␈α∂cognitive␈α∂analogue␈α∂of␈α⊂the␈α∂polyhedra
␈↓ α←␈↓prototypes.
␈↓ α←␈↓␈↓ β?But␈α∂is␈α∂it␈α∂reasonable␈α∂to␈α∂expect␈α∂that␈α∂such␈α∂structures␈α∂exist?␈α∂There␈α∞are
␈↓ α←␈↓several␈α
reasons␈α∞why␈α
the␈α∞idea␈α
seems␈α
plausible.␈α∞ Concepts␈α
in␈α∞human␈α
memory,
␈↓ α←␈↓for␈αinstance,␈αare␈αstructured␈αin␈αan␈αimplicit,␈αbut␈αhighly␈αcomplex␈αfashion.␈α
 They
␈↓ α←␈↓rarely␈αexist␈αindependently,␈αbut␈αrather␈αhave␈αa␈αnumber␈αof␈αcross␈αreferences␈αand
␈↓ α←␈↓interrelationships.␈α New␈α
knowledge␈αis␈α
typically␈α``filed␈α
away,''␈αrather␈α
than␈αjust
␈↓ α←␈↓deposited,␈α⊗and␈α⊗it␈α↔is␈α⊗clear␈α⊗when␈α⊗some␈α↔new␈α⊗item␈α⊗``doesn't␈α⊗fit''␈α↔into␈α⊗the
␈↓ α←␈↓established structure (see, e.g., [Norman75]).
␈↓ α←␈↓␈↓ β?There␈α
are␈α
also␈α∞suggestive␈α
regularities␈α
in␈α
the␈α∞set␈α
of␈α
rules␈α
in␈α∞the␈α
␈↓¬MYCIN␈↓
␈↓ α←␈↓system.␈α⊂ They␈α∂were␈α⊂first␈α∂noticed␈α⊂in␈α∂a␈α⊂naive␈α∂survey␈α⊂of␈α∂the␈α⊂knowledge␈α∂base
␈↓ α←␈↓when␈αthe␈αauthor␈αbegan␈α
working␈αwith␈αthe␈αsystem␈α
and␈αwere␈αlater␈αverified␈αby␈α
a
␈↓ α←␈↓simple␈αexperiment␈αwhich␈αdemonstrated␈αthat␈αthree␈αof␈αthe␈αclinicians␈αassociated
␈↓ α←␈↓with the project each, independently, detected the same regularities.␈↓
3␈↓
␈↓ α←␈↓␈↓ β?The␈αcognitive␈αmodels␈αwill␈αthus␈αbe␈αbased␈αon␈αuniformities␈αin␈αsubsets␈αof
␈↓ α←␈↓rules␈αand␈αwill␈α
look␈αlike␈αabstract␈α
descriptions␈αof␈αthose␈α
subsets,␈αbased␈αon␈α
a␈αset

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[3]␈αEach␈αof␈αthe␈αclinicians␈αwas␈αgiven␈αa␈αcomplete␈αset␈αof␈αrules␈αand␈αasked␈αto␈αsort
␈↓ α←␈↓them␈αinto␈α``similarity␈αclasses,''␈αsubsets␈αof␈αrules␈αthat␈αseemed␈αto␈αbelong␈αtogether.
␈↓ α←␈↓The␈α∂criteria␈α∞for␈α∂similarity␈α∂were␈α∞purposefully␈α∂left␈α∂up␈α∞to␈α∂each␈α∂clinician,␈α∞who
␈↓ α←␈↓was␈α∂asked␈α∂to␈α∞label␈α∂each␈α∂subset␈α∞to␈α∂explain␈α∂what␈α∞kind␈α∂of␈α∂rules␈α∂it␈α∞contained.
␈↓ α←␈↓The␈α
results␈α
strongly␈α
confirmed␈α
initial␈α
impressions--all␈α
the␈α
clinicians␈α
chose␈α
the
␈↓ α←␈↓same␈α
set␈αof␈α
similarity␈α
classes␈α(except␈α
that␈α
some␈αclasses␈α
had␈α
occasionally␈αbeen
␈↓ α←␈↓further␈α∞subdivided),␈α∂and␈α∞of␈α∞300␈α∂rules,␈α∞there␈α∞were␈α∂substantial␈α∞disagreements
␈↓ α←␈↓on the classification of only six.
␈↓ α←␈↓␈↓80    KNOWLEDGE ACQUISITION I␈↓ 
#5-3␈↓

␈↓"β␈↓ α←␈↓of␈αempirical␈αgeneralizations␈αabout␈αthem.␈α
 They␈αare␈αreferred␈αto␈αas␈α␈↓↓rule␈α
models␈↓.
␈↓ α←␈↓As␈αwill␈αbecome␈αclear,␈αthey␈αrepresent␈αa␈αsimple␈αform␈αof␈αconcept␈αformation--the
␈↓ α←␈↓concept␈α⊃of␈α⊃a␈α∩``typical''␈α⊃rule␈α⊃in␈α∩that␈α⊃subset.␈α⊃ They␈α∩will␈α⊃be␈α⊃used␈α∩to␈α⊃express
␈↓ α←␈↓expectations␈α∩about␈α∩the␈α∩expert's␈α∩new␈α∩rule␈α∩and␈α∩to␈α∩guide␈α∩processing␈α∩of␈α⊃the
␈↓ α←␈↓English text.
␈↓ α←␈↓␈↓ β?Rule models are composed of four parts (Fig. 5-1).

␈↓"β␈↓ α←␈↓	␈↓αEXAMPLES␈↓	␈↓ ∧␈the subset of rules which this model
␈↓"β␈↓ α←␈↓	␈↓ ∧␈describes

␈↓"β␈↓ α←␈↓	␈↓αDESCRIPTION␈↓	␈↓ ∧␈characterization of a "typical" member of
␈↓"β␈↓ α←␈↓	␈↓ ∧␈this subset
␈↓"β␈↓ α←␈↓	␈↓ ∧␈  characterization of the premise
␈↓"β␈↓ α←␈↓	␈↓ ∧␈  characterization of the action
␈↓"β␈↓ α←␈↓	␈↓ ∧␈    which attributes "typically" appear
␈↓"β␈↓ α←␈↓	␈↓ ∧␈    correlations of attributes

␈↓"β␈↓ α←␈↓	␈↓αMORE GENERAL␈↓	␈↓ ∧␈pointer to more general models
␈↓"β␈↓ α←␈↓	␈↓αMORE SPECIFIC␈↓	␈↓ ∧␈pointer to more specific models


␈↓"β␈↓ α←␈↓α␈↓ ∧jFig. 5-1.    Rule model structure.    

␈↓ α←␈↓They␈α∞contain,␈α∞first,␈α
a␈α∞list␈α∞of␈α
␈↓	EXAMPLES␈↓,␈α∞the␈α∞subset␈α
of␈α∞rules␈α∞from␈α∞which␈α
this
␈↓ α←␈↓model␈αwas␈αconstructed.␈α Next,␈αa␈α␈↓	DESCRIPTION␈↓␈αcharacterizes␈αa␈αtypical␈α
member
␈↓ α←␈↓of␈α⊃the␈α⊃subset.␈α⊃ Since␈α⊃we␈α⊃are␈α⊃dealing␈α⊃in␈α⊃this␈α⊃case␈α⊃with␈α⊃rules␈α⊃composed␈α⊂of
␈↓ α←␈↓premise-action␈α→pairs,␈α→the␈α→␈↓	DESCRIPTION␈↓␈α→currently␈α→implemented␈α_contains
␈↓ α←␈↓individual␈αcharacterizations␈αof␈αa␈αtypical␈αpremise␈αand␈αa␈αtypical␈αaction.␈α Then,
␈↓ α←␈↓since␈α∪the␈α∩current␈α∪representation␈α∪scheme␈α∩used␈α∪in␈α∪those␈α∩rules␈α∪is␈α∪based␈α∩on
␈↓ α←␈↓associative␈αtriples,␈α
we␈αhave␈α
implemented␈αthose␈α
characterizations␈αby␈α
indicating
␈↓ α←␈↓(a)␈α
the␈α
attributes␈αthat␈α
typically␈α
appear␈αin␈α
the␈α
premise␈α
(action)␈αof␈α
a␈α
rule␈αin␈α
this
␈↓ α←␈↓subset,␈α∃and␈α∃(b)␈α∃the␈α∃correlations␈α∃of␈α∃attributes␈α∃appearing␈α∃in␈α∃the␈α∃premise
␈↓ α←␈↓(action).
␈↓ α←␈↓␈↓ β?Note␈α∂that␈α∞the␈α∂central␈α∂idea␈α∞is␈α∂the␈α∂concept␈α∞of␈α∂␈↓↓characterizing␈α∂a␈α∞typical
␈↓ α←␈↓↓member␈α∂of␈α∂the␈α∂subset␈↓.␈α∂ Naturally,␈α∂that␈α∂characterization␈α∂would␈α∂look␈α∂different
␈↓ α←␈↓for␈α⊗subsets␈α⊗of␈α∃rules,␈α⊗procedures,␈α⊗theorems,␈α⊗etc.␈α∃ But␈α⊗the␈α⊗main␈α⊗idea␈α∃of
␈↓ α←␈↓characterization␈α∪is␈α∩widely␈α∪applicable␈α∩and␈α∪not␈α∩restricted␈α∪to␈α∪any␈α∩particular
␈↓ α←␈↓representational formalism.
␈↓ α←␈↓␈↓ β?The␈α∞two␈α∞other␈α∞parts␈α∞of␈α∞the␈α∞rule␈α∞model␈α∞are␈α∞pointers␈α∞to␈α∞other␈α
models
␈↓ α←␈↓describing␈α∪more␈α∪general␈α∪and␈α∪more␈α∩specific␈α∪subsets␈α∪of␈α∪rules.␈α∪ The␈α∪set␈α∩of
␈↓ α←␈↓models␈α
is␈α
organized␈α
into␈α
a␈α
number␈α∞of␈α
tree␈α
structures,␈α
Fig.␈α
5-2.␈α
 At␈α∞the␈α
root
␈↓ α←␈↓of␈α
each␈αtree␈α
is␈αthe␈α
model␈α
made␈αfrom␈α
all␈αthe␈α
rules␈α
that␈αconclude␈α
about␈αa␈α
given
␈↓ α←␈↓attribute␈α
(e.g.,␈αthe␈α
␈↓	CATEGORY␈↓␈αmodel);␈α
below␈αthis␈α
are␈αtwo␈α
models␈α
dealing␈αwith
␈↓ α←␈↓all␈α⊃affirmative␈α⊃and␈α⊃all␈α⊃negative␈α⊃rules␈α⊃(e.g.,␈α⊃the␈α⊃␈↓	CATEGORY-IS␈↓␈α∩model);␈α⊃and
␈↓ α←␈↓below␈α
this␈αare␈α
models␈αdealing␈α
with␈α
rules␈αthat␈α
affirm␈αor␈α
deny␈α
specific␈αvalues
␈↓ α←␈↓of the attribute.
␈↓ α←␈↓␈↓5-3␈↓ πbRULE MODEL OVERVIEW    81␈↓


␈↓"β␈↓ α←␈↓∧                           <attr>
␈↓"β␈↓ α←␈↓∧                        ≤          ≥
␈↓"β␈↓ α←␈↓∧                     ≤                ≥
␈↓"β␈↓ α←␈↓∧                  ≤                      ≥
␈↓"β␈↓ α←␈↓∧               ≤                            ≥
␈↓"β␈↓ α←␈↓∧        <attr>-is                     <attr>-isnot
␈↓"β␈↓ α←␈↓∧        /       \                      /          \
␈↓"β␈↓ α←␈↓∧       /         \                    /            \
␈↓"β␈↓ α←␈↓∧      /           \                  /              \

␈↓"β␈↓ α←␈↓∧<attr>-is-X   <attr>-is-Y    <attr>-isnot-X   <attr>-isnot-Y


␈↓"β␈↓ α←␈↓α␈↓ ∧9Fig. 5-2.    Organization of rule models.    

␈↓ α←␈↓␈↓ β?For␈αany␈αgiven␈αattribute,␈αsome␈αof␈αthe␈αbranches␈αmay␈αnot␈αexist␈αsince␈αwe
␈↓ α←␈↓require␈α(empirically)␈αthat␈αthere␈αmust␈αbe␈αat␈αleast␈αtwo␈αrelevant␈αrules␈αbefore␈αthe
␈↓ α←␈↓corresponding␈αmodel␈αis␈αcreated␈α(which␈αwas␈αwhy␈αthere␈αwas␈αno␈α␈↓	CATEGORY-IS-
␈↓ α←␈↓	ENTEROBACTERIACEAE␈↓ model in the current system).
␈↓ α←␈↓␈↓ β?The␈α∂fact␈α∂that␈α∂the␈α∂models␈α∂are␈α∂organized␈α∂around␈α∂the␈α∂contents␈α∂of␈α∞the
␈↓ α←␈↓action␈α∂part␈α∂assumes␈α⊂that␈α∂rules␈α∂which␈α∂conclude␈α⊂about␈α∂the␈α∂same␈α⊂thing␈α∂have
␈↓ α←␈↓useful␈αsimilarities␈αin␈αtheir␈αpremises.␈α This␈αassumption␈αwas␈αmade␈αplausible␈αby
␈↓ α←␈↓several␈α∞considerations.␈α∂ To␈α∞begin␈α∂with,␈α∞this␈α∂organization␈α∞paralleled␈α∂the␈α∞one
␈↓ α←␈↓used␈α∞by␈α∂the␈α∞clinicians␈α∂when␈α∞they␈α∂were␈α∞asked␈α∞to␈α∂sort␈α∞the␈α∂rules.␈α∞ It␈α∂was␈α∞also
␈↓ α←␈↓hoped␈αthat␈αthe␈αmodels␈αwould␈αbe␈αmore␈αthan␈αjust␈αa␈αdescription␈αof␈αa␈α
subset␈αof
␈↓ α←␈↓rules,␈α
that␈α∞they␈α
might␈α
also␈α∞suggest␈α
things␈α
about␈α∞the␈α
reasoning␈α∞process␈α
itself.
␈↓ α←␈↓It␈α⊃is␈α∩possible,␈α⊃for␈α⊃instance,␈α∩that␈α⊃simple␈α⊃correlations␈α∩among␈α⊃terms␈α∩in␈α⊃rules
␈↓ α←␈↓might␈αactually␈αreflect␈αa␈αvalid␈αgeneralization␈αabout␈αthe␈αway␈αan␈αexpert␈αreasons
␈↓ α←␈↓in␈α⊃this␈α⊃domain.␈α⊂ In␈α⊃the␈α⊃current␈α⊂implementation,␈α⊃therefore,␈α⊃the␈α⊃models␈α⊂are
␈↓ α←␈↓organized␈α
around␈α
the␈α
action␈α
part␈α
of␈α
the␈α
rule;␈α
but,␈α
as␈α
discussed␈α∞below,␈α
other
␈↓ α←␈↓organizations are possible.
␈↓ α←␈↓␈↓ β?Rather␈α
than␈αbeing␈α
hand-tooled,␈α
the␈αmodels␈α
are␈α
assembled␈αby␈α
␈↓¬TEIRESIAS␈↓
␈↓ α←␈↓on␈αthe␈αbasis␈αof␈αthe␈αcurrent␈αcontents␈αof␈αthe␈αknowledge␈αbase,␈αin␈αwhat␈αamounts
␈↓ α←␈↓to␈α∀a␈α∪very␈α∀simple␈α∪(i.e.,␈α∀statistical)␈α∪form␈α∀of␈α∪concept␈α∀formation.␈α∀ Thus,␈α∪the
␈↓ α←␈↓combination␈α
of␈α␈↓¬TEIRESIAS␈↓␈α
and␈αthe␈α
performance␈α
program␈αpresents␈α
a␈αsystem␈α
that
␈↓ α←␈↓has a model of its own knowledge, a model it forms by itself.

␈↓ α←␈↓␈↓α5-3-3    Rule models:  Example␈↓
␈↓ α←␈↓␈↓ β?Shown␈αbelow␈αis␈αan␈αexample␈αof␈αa␈αrule␈αmodel␈αfrom␈αthe␈αcurrent␈αsystem.
␈↓ α←␈↓This␈α≥particular␈α≥model␈α≥describes␈α≥the␈α≥subset␈α≥of␈α≥rules␈α≡that␈α≥conclude
␈↓ α←␈↓affirmatively␈αabout␈α
the␈αcategory␈α
of␈αan␈α
organism␈αand␈α
is␈αthe␈α
one␈αthat␈αwas␈α
used
␈↓ α←␈↓by␈α
the␈α
system␈α
in␈αthe␈α
trace␈α
shown␈α
earlier.␈α
 Each␈αof␈α
the␈α
pieces␈α
of␈α
the␈αmodel␈α
will
␈↓ α←␈↓be␈α⊂described␈α⊂in␈α⊂some␈α⊂detail,␈α⊂to␈α∂give␈α⊂the␈α⊂reader␈α⊂a␈α⊂feeling␈α⊂for␈α⊂the␈α∂concepts
␈↓ α←␈↓involved.␈α The␈αusual␈αwarnings␈αapply,␈αhowever: ␈αIt␈αis␈αthe␈αglobal␈αconcepts␈αthat
␈↓ α←␈↓are␈α∪important␈α∪here,␈α∩rather␈α∪than␈α∪the␈α∪precise␈α∩format␈α∪or␈α∪location␈α∪of␈α∩every
␈↓ α←␈↓␈↓82    KNOWLEDGE ACQUISITION I␈↓ 
#5-3␈↓

␈↓"β␈↓ α←␈↓parenthesis.␈α∩ The␈α∩models␈α∪are␈α∩also␈α∩fairly␈α∩compact,␈α∪in␈α∩the␈α∩sense␈α∪that␈α∩they
␈↓ α←␈↓contain␈α∩a␈α∩great␈α∪deal␈α∩of␈α∩information.␈α∪ We␈α∩will␈α∩see␈α∩that␈α∪all␈α∩of␈α∩it␈α∪is␈α∩used
␈↓ α←␈↓eventually.

␈↓"β␈↓ α←␈↓¬    ␈↓&CATEGORY-IS␈↓)αβ

␈↓"β␈↓ α←␈↓¬    EXAMPLES   ((RULE116  .33)
␈↓"β␈↓ α←␈↓¬                (RULE050  .70)
␈↓"β␈↓ α←␈↓¬                (RULE037  .80)
␈↓"β␈↓ α←␈↓¬                (RULE095  .90)
␈↓"β␈↓ α←␈↓¬                (RULE152 1.0)
␈↓"β␈↓ α←␈↓¬                (RULE140 1.0))

␈↓"β␈↓ α←␈↓¬    DESCRIPTION
␈↓"β␈↓ α←␈↓¬      PREMISE  ((GRAM SAME NOTSAME 3.83)
␈↓"β␈↓ α←␈↓¬                (MORPH SAME NOTSAME 3.83)
␈↓"β␈↓ α←␈↓¬                ((GRAM SAME) (MORPH SAME) 3.83)
␈↓"β␈↓ α←␈↓¬                ((MORPH SAME) (GRAM SAME) 3.83)
␈↓"β␈↓ α←␈↓¬                ((AIR SAME) (NOSOCOMIAL NOTSAME SAME) (MORPH SAME)(GRAM SAME) 1.50)
␈↓"β␈↓ α←␈↓¬                ((NOSOCOMIAL NOTSAME SAME) (AIR SAME) (MORPH SAME)(GRAM SAME) 1.50)
␈↓"β␈↓ α←␈↓¬                ((INFECTION SAME) (SITE MEMBF SAME) 1.23)
␈↓"β␈↓ α←␈↓¬                ((SITE MEMBF SAME) (INFECTION SAME) (PORTAL SAME) 1.23))

␈↓"β␈↓ α←␈↓¬      ACTION   ((CATEGORY CONCLUDE 4.73)
␈↓"β␈↓ α←␈↓¬                (IDENT CONCLUDE 4.05)
␈↓"β␈↓ α←␈↓¬                ((CATEGORY CONCLUDE) (IDENT CONCLUDE) 4.73))

␈↓"β␈↓ α←␈↓¬    MORE-GENL  (CATEGORY)

␈↓"β␈↓ α←␈↓¬    MORE-SPEC   NIL

␈↓ α←␈↓α␈↓ β≠Fig.␈α∞5-3.    Rule␈α∞model␈α∂for␈α∞rules␈α∞concluding␈α∂affirmatively␈α∞about
␈↓ α←␈↓α␈↓ β≠category.    

␈↓ α←␈↓␈↓ β?This␈α
model␈α
was␈α
formed␈α
from␈α
the␈α
six␈α
rules␈α
listed␈α
in␈α
the␈α␈↓	EXAMPLES␈↓␈α
part
␈↓ α←␈↓of␈α∞Fig.␈α∞5-3.␈α∞ (The␈α
numbers␈α∞between␈α∞zero␈α∞and␈α
one␈α∞associated␈α∞with␈α∞the␈α
rules
␈↓ α←␈↓are␈α
the␈α
relevant␈α∞certainty␈α
factors␈α
from␈α
the␈α∞conclusions;␈α
these␈α
are␈α∞used␈α
when
␈↓ α←␈↓updating the models in response to changes in the knowledge base.)
␈↓ α←␈↓␈↓ β?The␈α⊃␈↓	DESCRIPTION␈↓s␈α⊃of␈α⊃the␈α⊃premise␈α⊃and␈α⊃action␈α⊃indicate␈α⊃regularities
␈↓ α←␈↓found␈α
in␈α
the␈α
subset␈αof␈α
rules␈α
listed␈α
in␈α␈↓	EXAMPLES␈↓.␈α
 The␈α
two␈α
types␈αof␈α
description
␈↓ α←␈↓noted␈α⊃above␈α⊂are␈α⊃each␈α⊂encoded␈α⊃in␈α⊂their␈α⊃own␈α⊂format: ␈α⊃There␈α⊃are␈α⊂``singlets''
␈↓ α←␈↓indicating␈α≤which␈α≠attributes␈α≤typically␈α≠appear␈α≤and␈α≤``ntuples''␈α≠indicating
␈↓ α←␈↓(empirical)␈α⊃correlations␈α⊂of␈α⊃attributes.␈α⊃ For␈α⊂instance,␈α⊃the␈α⊂first␈α⊃singlet␈α⊃in␈α⊂the
␈↓ α←␈↓premise description above

␈↓"β␈↓ α←␈↓	␈↓ ¬↔(GRAM SAME NOTSAME 3.83)

␈↓ α←␈↓indicates␈αthat␈αgramstain␈α(␈↓	GRAM␈↓)␈αtypically␈αappears␈αand␈αis␈αoften␈αassociated␈αwith
␈↓ α←␈↓the functions ␈↓	SAME␈↓ and ␈↓	NOTSAME␈↓.  (The 3.83 is explained below.)
␈↓ α←␈↓␈↓ β?The␈α
third␈α
item␈α
in␈α
the␈α
premise␈α
description␈α
of␈α
Fig.␈α
5-3␈α
is␈α
an␈αexample
␈↓ α←␈↓of an ntuple:
␈↓"β␈↓ α←␈↓␈↓5-3␈↓ πbRULE MODEL OVERVIEW    83␈↓

␈↓"β␈↓ α←␈↓	␈↓ ∧g((GRAM SAME)(MORPH SAME) 3.83)

␈↓ α←␈↓It␈α
indicates␈αthat␈α
whenever␈αgramstain␈α
appears␈αin␈α
a␈αrule␈α
premise,␈αmorphology
␈↓ α←␈↓(␈↓	MORPH␈↓) is typically found as well.
␈↓ α←␈↓␈↓ β?A␈α
collection␈α
of␈α
this␈α
sort␈α
of␈α
information␈α
can␈α
provide␈α
a␈α
fairly␈αdetailed
␈↓ α←␈↓picture␈α
of␈α∞the␈α
``typical''␈α∞appearance␈α
of␈α∞a␈α
rule␈α∞from␈α
this␈α∞subset.␈α
 In␈α∞this␈α
case,
␈↓ α←␈↓rules␈α
that␈α
conclude␈α
affirmatively␈α
about␈α
category␈α
``typically''␈αmention␈α
gramstain
␈↓ α←␈↓and␈αmorphology;␈α
when␈αone␈αof␈α
these␈αappears,␈αthe␈α
other␈αtends␈αto␈α
appear␈αalso,
␈↓ α←␈↓and so on.
␈↓ α←␈↓␈↓ β?Finally,␈αthe␈α␈↓	MORE-SPEC␈↓␈αand␈α␈↓	MORE-GENL␈↓␈αpointers␈αindicate␈αthat␈αthere␈αis
␈↓ α←␈↓no␈αrule␈αmodel␈αmore␈αspecific␈αthan␈αthis␈αone␈αin␈αthe␈αcurrent␈αsystem␈αand␈αthat␈αthe
␈↓ α←␈↓␈↓	CATEGORY␈↓ model is more general.

␈↓ α←␈↓␈↓αDetails of the ␈↓	DESCRIPTION␈↓α implementation␈↓	    
␈↓ α←␈↓␈↓ β?Each singlet is of the form

␈↓"β␈↓ α←␈↓	␈↓ ∧α( <attrib> <predicate-function>␈↓
+␈↓	 <cf-sum> )

␈↓ α←␈↓where␈αthe␈αsuperscript␈α``+''␈αindicates␈αone␈αor␈αmore␈αof␈αinstantiations␈αof␈αthe␈αitem.
␈↓ α←␈↓A␈αsinglet␈αof␈αthis␈α
sort␈αwill␈αbe␈αconstructed␈αfor␈α
any␈αattribute␈αthat␈αappears␈αin␈α
the
␈↓ α←␈↓premises␈α∀of␈α∀at␈α∀least␈α∪30%␈α∀of␈α∀the␈α∀rules␈α∪in␈α∀the␈α∀subset.␈↓
4␈↓␈α∀Enough␈α∪predicate
␈↓ α←␈↓functions␈αare␈αthen␈αsupplied␈αto␈αaccount␈αfor␈αat␈αleast␈α75%␈αof␈αthe␈αappearances␈αof
␈↓ α←␈↓the␈αattribute.␈α Thus␈αthe␈αexample␈α
cited␈αabove␈αmeans␈αthat␈α(a)␈αgramstain␈α
occurs
␈↓ α←␈↓in␈α
a␈α∞premise␈α
clause␈α∞of␈α
at␈α∞least␈α
30%␈α∞of␈α
the␈α∞rules␈α
listed␈α∞in␈α
␈↓	EXAMPLES␈↓,␈α∞and␈α
(b)
␈↓ α←␈↓75%␈αof␈αthose␈αclauses␈αcontaining␈αgramstain␈αuse␈αeither␈α␈↓	SAME␈↓␈αor␈α␈↓	NOTSAME␈↓␈αas␈αthe
␈↓ α←␈↓predicate function.
␈↓ α←␈↓␈↓ β?The␈α∞3.83␈α∞at␈α∞the␈α∞end␈α∞is␈α∞the␈α∞sum␈α∞of␈α∞the␈α∞certainty␈α∞factors␈α∞of␈α∞the␈α∞rules
␈↓ α←␈↓that␈α∂had␈α∂premise␈α∂clauses␈α⊂mentioning␈α∂gramstain.␈α∂ As␈α∂will␈α∂become␈α⊂clear,␈α∂the
␈↓ α←␈↓␈↓	DESCRIPTION␈↓s␈α∞contained␈α∞in␈α
the␈α∞rule␈α∞models␈α
provide␈α∞␈↓¬TEIRESIAS␈↓␈α∞with␈α
``advice''
␈↓ α←␈↓about␈α∞what␈α∂to␈α∞expect␈α∂to␈α∞find␈α∂in␈α∞a␈α∞newly␈α∂acquired␈α∞rule;␈α∂in␈α∞these␈α∂terms,␈α∞the
␈↓ α←␈↓numbers␈α∂become␈α⊂an␈α∂indication␈α⊂of␈α∂the␈α⊂``strength''␈α∂of␈α⊂this␈α∂advice.␈α⊂ While␈α∂its
␈↓ α←␈↓origin␈α
is␈α∞purely␈α
empirical,␈α
there␈α∞are␈α
reasons␈α
why␈α∞it␈α
appears␈α
to␈α∞be␈α
plausible.
␈↓ α←␈↓First,␈α∂by␈α∂summing␈α∞up␈α∂CFs,␈α∂the␈α∞strongest␈α∂advice␈α∂will␈α∞arise␈α∂from␈α∂the␈α∞factors
␈↓ α←␈↓that␈αappear␈α
the␈αmost␈αoften.␈α
 Since␈αthe␈αmodel␈α
is␈αsupposed␈αto␈α
characterize␈αthe
␈↓ α←␈↓typical␈α↔content␈α↔of␈α_a␈α↔rule,␈α↔a␈α↔dependency␈α_of␈α↔strength␈α↔on␈α_frequency␈α↔of
␈↓ α←␈↓appearance␈αseems␈α
reasonable.␈α Second,␈αrules␈α
with␈αhigh␈αCFs␈α
tend␈αto␈α
be␈αthose
␈↓ α←␈↓that␈α⊗are␈α⊗the␈α⊗soundest␈α⊗and␈α⊗perhaps␈α⊗the␈α⊗closest␈α⊗to␈α⊗being␈α⊗based␈α⊗on␈α⊗an
␈↓ α←␈↓understanding␈α∩of␈α∪underlying␈α∩processes␈α∩rather␈α∪than␈α∩simply␈α∪on␈α∩experience.
␈↓ α←␈↓Thus,␈α∃given␈α∃equal␈α∃frequency␈α⊗of␈α∃appearance,␈α∃advice␈α∃arising␈α⊗from␈α∃more
␈↓ α←␈↓definite␈α
rules␈α
will␈α
be␈α
stronger.␈α
 Finally,␈αit␈α
appears␈α
to␈α
give␈α
a␈α
satisfying␈αrange
␈↓ α←␈↓of␈α
strengths;␈α
in␈α
the␈α∞current␈α
system,␈α
the␈α
largest␈α
such␈α∞value␈α
is␈α
close␈α
to␈α∞21,␈α
the
␈↓ α←␈↓smallest is 0.7.

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[4]␈α∩This␈α∩and␈α∩all␈α∩other␈α∪thresholds␈α∩cited␈α∩are␈α∩purely␈α∩empirical.␈α∪They␈α∩were
␈↓ α←␈↓chosen␈α⊂solely␈α⊂on␈α⊂the␈α⊂basis␈α⊂of␈α⊂producing␈α⊂results␈α⊂that␈α⊂were␈α⊂similar␈α⊂to␈α⊂those
␈↓ α←␈↓suggested by the rule sorting experiment performed with the clinicians.
␈↓ α←␈↓␈↓84    KNOWLEDGE ACQUISITION I␈↓ 
#5-3␈↓

␈↓"β␈↓ α←␈↓␈↓ β?To␈α
review,␈α
the␈α
first␈αsinglet␈α
in␈α
the␈α
rule␈α
model␈αof␈α
Fig.␈α
5-3␈α
means␈αthat␈α
at
␈↓ α←␈↓least␈α∂30%␈α⊂of␈α∂the␈α∂rules␈α⊂in␈α∂␈↓	EXAMPLES␈↓␈α∂had␈α⊂a␈α∂premise␈α∂clause␈α⊂containing␈α∂gram
␈↓ α←␈↓stain,␈α∂that␈α∂at␈α∂least␈α∂75%␈α∂of␈α∂those␈α∂clauses␈α∂used␈α∂either␈α∂␈↓	SAME␈↓␈α∂or␈α∂␈↓	NOTSAME␈↓␈α∂and,
␈↓ α←␈↓finally, that the sum of the CFs of those rules was 3.83.
␈↓ α←␈↓␈↓ β?The general form of the ntuples is

␈↓"β␈↓ α←␈↓	␈↓ βm( (<attrib> <predicate-function>␈↓
+␈↓	)␈↓
+␈↓	 <cf-sum> )

␈↓ α←␈↓and the first example in this model is

␈↓"β␈↓ α←␈↓	␈↓ ∧←((GRAM SAME) (MORPH SAME) 3.83)

␈↓ α←␈↓This␈α∂ntuple␈α⊂means␈α∂that␈α∂whenever␈α⊂gram␈α∂stain␈α∂appeared␈α⊂in␈α∂a␈α⊂rule␈α∂premise,
␈↓ α←␈↓morphology␈α(␈↓	MORPH␈↓)␈αalso␈αappeared␈αat␈αleast␈α80%␈αof␈αthe␈αtime.␈α The␈αimplication
␈↓ α←␈↓is␈αunidirectional--if␈αit␈α
is␈αalso␈αtrue␈α
that␈αwhenever␈αmorphology␈α
appears,␈αgram
␈↓ α←␈↓stain␈α∂also␈α∂appears␈α∂(at␈α⊂least␈α∂80%␈α∂of␈α∂the␈α⊂time),␈α∂then␈α∂there␈α∂would␈α⊂be␈α∂another
␈↓ α←␈↓ntuple␈α∞expressing␈α
this␈α∞fact␈α
(as␈α∞there␈α
is␈α∞in␈α
this␈α∞case).␈↓
5␈↓␈α
All␈α∞this␈α∞generalizes␈α
to
␈↓ α←␈↓more␈α∪than␈α∪two␈α∩attributes␈α∪so␈α∪that,␈α∩for␈α∪example,␈α∪the␈α∩fifth␈α∪element␈α∪of␈α∩the
␈↓ α←␈↓␈↓	DESCRIPTION␈↓␈α⊃indicates␈α⊃that␈α⊃the␈α⊃presence␈α⊃of␈α⊃aerobicity␈α⊃(␈↓	AIR␈↓)␈α⊃indicates␈α⊂the
␈↓ α←␈↓likely presence of several other attributes.

␈↓ α←␈↓␈↓α5-3-4    Rule models as concept formation␈↓
␈↓ α←␈↓␈↓ β?As␈αnoted,␈α
the␈αrule␈αmodels␈α
are␈αconstructed␈αby␈α
␈↓¬TEIRESIAS␈↓␈αfrom␈α
the␈αrules
␈↓ α←␈↓in␈α∞the␈α∞performance␈α∞program's␈α
knowledge␈α∞base--the␈α∞description␈α∞of␈α∞the␈α
rule-
␈↓ α←␈↓model␈α
components␈αgiven␈α
in␈αthe␈α
previous␈αsection␈α
is␈αan␈α
informal␈αspecification
␈↓ α←␈↓of the algorithm used.
␈↓ α←␈↓␈↓ β?Despite␈α_this␈α_automated␈α↔construction,␈α_however,␈α_there␈α_are␈α↔several
␈↓ α←␈↓reasons␈α⊂why␈α⊂the␈α⊂rule␈α⊂models␈α⊂are␈α⊂only␈α⊂a␈α⊂weak␈α⊂form␈α⊂of␈α⊃concept␈α⊂formation.
␈↓ α←␈↓First,␈α
and␈α
perhaps␈α
most␈αimportant,␈α
the␈α
concepts␈α
are␈αexpressed␈α
in␈α
terms␈α
of␈αa
␈↓ α←␈↓predefined␈αset␈α
of␈αpatterns: ␈α
the␈αsinglets␈α
and␈αntuples,␈α
both␈αused␈α
to␈αdescribe␈α
the
␈↓ α←␈↓fixed␈αsubsets␈α
of␈αrules␈α
described␈αin␈αthe␈α
previous␈αsection.␈α
 While␈αit␈α
would␈αnot
␈↓ α←␈↓be␈α⊂difficult␈α⊂to␈α⊂add␈α⊂the␈α⊂ability␈α⊂to␈α⊂detect␈α⊂other␈α⊂types␈α⊂of␈α⊂patterns␈α⊃in␈α⊂subsets
␈↓ α←␈↓organized␈α→along␈α→different␈α→criteria,␈α→this␈α→would␈α→not␈α→confront␈α→the␈α→more
␈↓ α←␈↓fundamental␈α⊂problem␈α⊂of␈α⊂concept␈α⊂formation: ␈α⊂the␈α⊂judicious␈α⊂selection␈α⊂of␈α⊂the
␈↓ α←␈↓``proper''␈α
organization␈α
of␈α
the␈α
examples␈α
to␈α
detect␈α
``important''␈α
regularities␈α
and
␈↓ α←␈↓the ability to discover unpredicted patterns.
␈↓ α←␈↓␈↓ β?In␈α∞addition,␈α
the␈α∞current␈α∞implementation␈α
commits␈α∞what␈α∞Winston␈α
calls
␈↓ α←␈↓the␈α
``fallacy␈α
that␈α
frequent␈α
appearance␈α
means␈α
importance''␈α
[Winston70].␈α
 While
␈↓ α←␈↓the␈α∀strength␈α∪of␈α∀each␈α∪piece␈α∀of␈α∪advice␈α∀in␈α∪the␈α∀models␈α∪does␈α∀provide␈α∪some

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[5]␈αThe␈αinterpretation␈αof␈αthe␈α
<predicate-function>␈αand␈α<cf-sum>␈αis␈αthe␈α
same,
␈↓ α←␈↓except␈αthat␈αthe␈α
<predicate-function>s␈αare␈αtaken␈α
only␈αfrom␈αthose␈α
instances␈αin
␈↓ α←␈↓which␈αthe␈α<attrib>s␈αactually␈αdid␈αappear␈αtogether.␈α This␈αaccounts␈αfor␈αthe␈αfact
␈↓ α←␈↓that␈α⊂␈↓	NOTSAME␈↓␈α⊃does␈α⊂not␈α⊃appear.␈α⊂Evidently,␈α⊂when␈α⊃gram␈α⊂and␈α⊃morph␈α⊂appear
␈↓ α←␈↓together, they tend to use just ␈↓	SAME␈↓.
␈↓ α←␈↓␈↓5-3␈↓ πbRULE MODEL OVERVIEW    85␈↓

␈↓"β␈↓ α←␈↓measure␈α∃of␈α⊗differential␈α∃importance␈α⊗(similar␈α∃to␈α⊗the␈α∃``must-have''␈α⊗link␈α∃in
␈↓ α←␈↓[Winston70]),␈α⊃that␈α⊃measure␈α⊃is␈α∩acquired␈α⊃purely␈α⊃on␈α⊃the␈α⊃basis␈α∩of␈α⊃frequency.
␈↓ α←␈↓This␈α
strongly␈αlimits␈α
the␈αrange␈α
of␈αconcepts␈α
that␈αcan␈α
be␈αdefined: ␈α
There␈α
is␈αno
␈↓ α←␈↓way,␈αfor␈α
instance,␈αto␈αinfer␈α
a␈αlink␈αequivalent␈α
to␈αa␈α``must-not-have''␈α
link.␈α Nor
␈↓ α←␈↓is␈αthere␈αany␈αnotion␈αof␈α␈↓↓why␈↓␈αa␈αparticular␈αcollection␈αof␈αattributes␈αtend␈αto␈αappear
␈↓ α←␈↓together.␈α⊗ The␈α⊗current␈α⊗approach␈α∃can␈α⊗not␈α⊗distinguish␈α⊗between␈α∃statistical
␈↓ α←␈↓accidents␈α⊗and␈α⊗a␈α↔meaningful␈α⊗correlation.␈α⊗In␈α↔the␈α⊗absence␈α⊗of␈α↔a␈α⊗carefully
␈↓ α←␈↓constructed␈α⊂set␈α⊂of␈α⊃teaching␈α⊂examples␈α⊂(in␈α⊂particular,␈α⊃there␈α⊂are␈α⊂none␈α⊃of␈α⊂the
␈↓ α←␈↓``near␈α⊃misses''␈α⊃Winston␈α⊃uses␈α⊃to␈α⊂great␈α⊃advantage),␈α⊃the␈α⊃system␈α⊃uses␈α⊃the␈α⊂only
␈↓ α←␈↓information␈α⊃it␈α⊃has␈α⊃and␈α⊃assumes␈α⊃that␈α⊂all␈α⊃rules␈α⊃in␈α⊃the␈α⊃knowledge␈α⊃base␈α⊂are
␈↓ α←␈↓equally good instances.
␈↓ α←␈↓␈↓ β?Finally,␈αthe␈αpattern␈αdetection␈αroutines␈αall␈αuse␈αan␈αexhaustive␈αsearch␈αof
␈↓ α←␈↓the␈α∪knowledge␈α∪base.␈α∪ This␈α∀is␈α∪not␈α∪particularly␈α∪sophisticated,␈α∪and␈α∀a␈α∪more
␈↓ α←␈↓powerful approach to concept formation should be more efficient.␈↓
6␈↓
␈↓ α←␈↓␈↓ β?Despite␈αthe␈αsimplicity␈αof␈αthe␈αtechniques␈αused,␈αthe␈αrule␈αmodels␈αpresent
␈↓ α←␈↓an␈α⊂adequate␈α⊂picture␈α⊂of␈α∂the␈α⊂concept␈α⊂of␈α⊂a␈α∂``typical''␈α⊂rule.␈α⊂ There␈α⊂are␈α∂several
␈↓ α←␈↓reasons␈α
why␈α
this␈α
is␈α∞so.␈α
First,␈α
the␈α
concept␈α
is␈α∞ill-defined␈α
to␈α
begin␈α
with␈α∞and␈α
is
␈↓ α←␈↓being␈αformed␈αfrom␈αa␈αvery␈αlimited␈αset␈αof␈αdata.␈α This␈αimplies␈αthat␈αit␈αmight␈αnot
␈↓ α←␈↓be␈αwise␈α
to␈αdevote␈αextensive␈α
amounts␈αof␈α
computation␈αto␈αsophisticated␈α
concept-
␈↓ α←␈↓formation␈α∩routines.␈α∩ Second,␈α∩the␈α⊃fundamental␈α∩task␈α∩here␈α∩is␈α⊃comprehension
␈↓ α←␈↓rather␈α
than␈α
concept␈α
formation,␈α
and␈α
the␈αmodels␈α
are␈α
not␈α
the␈α
end␈α
product␈αof␈α
the
␈↓ α←␈↓program.␈α∂ The␈α∂models␈α⊂are␈α∂instead␈α∂used␈α∂later␈α⊂in␈α∂a␈α∂support␈α∂role␈α⊂to␈α∂provide
␈↓ α←␈↓more␈α⊂sophisticated␈α⊂comprehension.␈α⊃ Finally,␈α⊂exhaustive␈α⊂search␈α⊃for␈α⊂patterns
␈↓ α←␈↓does␈α⊂not␈α⊂present␈α⊂significant␈α⊂problems␈α⊂because␈α⊂the␈α⊂knowledge␈α⊂base␈α⊂is␈α∂built
␈↓ α←␈↓incrementally,␈α⊂hence␈α⊃rule␈α⊂model␈α⊂computation␈α⊃can␈α⊂proceed␈α⊂in␈α⊃parallel.␈α⊂ (As
␈↓ α←␈↓long␈α≠as␈α≠intermediate␈α≠results␈α≠are␈α≠kept␈α≠for␈α≠reference,␈α≠the␈α~incremental
␈↓ α←␈↓computation␈αis␈αnot␈α
time␈αconsuming.) ␈αImprovements␈αcould␈α
still␈αbe␈αmade␈αin␈α
all
␈↓ α←␈↓of␈αthese␈αareas,␈αbut␈α
the␈αpoint␈αhere␈αis␈α
simply␈αthat␈αvery␈αsophisticated␈α
techniques
␈↓ α←␈↓are not required for adequate performance.

␈↓ α←␈↓␈↓α5-3-5    Implications␈↓
␈↓ α←␈↓␈↓ β?A␈α∩number␈α∩of␈α∩implications␈α∩follow␈α∩from␈α∩the␈α∩design␈α∩and␈α∩automated
␈↓ α←␈↓construction␈αof␈αthe␈αmodels.␈α First,␈αthe␈αuse␈αof␈αthe␈αsinglets␈αand␈αntuples␈αis␈αreally
␈↓ α←␈↓only␈α∩one␈α∩instance␈α∩of␈α∩a␈α∩more␈α∩general␈α∩idea.␈α∩ In␈α∩fact,␈α∩any␈α∩kind␈α∩of␈α∩pattern
␈↓ α←␈↓involving␈α
any␈α
of␈α
the␈αcomponents␈α
of␈α
a␈α
rule␈αcould␈α
be␈α
used.␈α
 While␈α
they␈αwere
␈↓ α←␈↓not␈α∀included␈α∀(for␈α∪the␈α∀sake␈α∀of␈α∪simplicity),␈α∀patterns␈α∀in␈α∪the␈α∀values␈α∀of␈α∪the
␈↓ α←␈↓attributes␈α#or␈α"certainty␈α#factors␈α"might␈α#also␈α"have␈α#been␈α"considered.
␈↓ α←␈↓Straightforward␈α∪extensions␈α∪to␈α∪the␈α∩current␈α∪implementation␈α∪would␈α∪make␈α∩it
␈↓ α←␈↓possible␈α∞to␈α∞look␈α∞for␈α∞any␈α∞specified␈α∞pattern␈α∞involving␈α∞any␈α∞of␈α∂the␈α∞components
␈↓ α←␈↓and␈α
to␈α
use␈α
this␈α
in␈α
much␈α
the␈α
same␈αway␈α
as␈α
a␈α
picture␈α
of␈α
what␈α
to␈α
expect␈α
in␈αa␈α
rule
␈↓ α←␈↓of this subset.
␈↓ α←␈↓␈↓ β?As␈αnoted,␈αthe␈αrule␈αmodels␈αare␈αassembled␈αby␈αthe␈αsystem␈αfrom␈αthe␈αrules

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[6]␈α∪A␈α∀recent␈α∪report␈α∀[Hayes-Roth76]␈α∪discusses␈α∀more␈α∪formal␈α∀and␈α∪efficient
␈↓ α←␈↓methods for a similar sort of concept formation task.
␈↓ α←␈↓␈↓86    KNOWLEDGE ACQUISITION I␈↓ 
#5-3␈↓

␈↓"β␈↓ α←␈↓in␈α∪the␈α∪knowledge␈α∪base.␈α∪ This␈α∪automatic␈α∪generation␈α∪and␈α∪updating␈α∪has␈α∪a
␈↓ α←␈↓number of advantages.
␈↓ α←␈↓␈↓ β?In␈α
our␈αdiscussion␈α
of␈αFalk's␈α
original␈αconcept␈α
of␈αmodels,␈α
for␈αinstance,␈α
we
␈↓ α←␈↓noted␈αboth␈α
the␈αpotential␈α
disadvantage␈αof␈α
a␈αfixed␈α
set␈αof␈α
models␈αand␈αthe␈α
utility
␈↓ α←␈↓of␈α⊃being␈α⊃able␈α⊃to␈α⊃add␈α⊃new␈α⊃models␈α⊃easily.␈α⊃ The␈α⊃system␈α⊃described␈α⊃here␈α⊂has
␈↓ α←␈↓avoided␈α⊃the␈α⊂disadvantage␈α⊃of␈α⊃a␈α⊂fixed␈α⊃set␈α⊂and␈α⊃has␈α⊃the␈α⊂ability␈α⊃to␈α⊃add␈α⊂new
␈↓ α←␈↓models without difficulty.  In addition:

␈↓ α←␈↓␈↓ ββ(1)␈↓ β?The␈α∂expert␈α∞need␈α∂never␈α∞concern␈α∂himself␈α∞with␈α∂adding␈α∞models,
␈↓ α←␈↓␈↓ β?since␈α␈↓¬TEIRESIAS␈↓␈α
takes␈αcare␈αof␈α
all␈αthe␈α
work␈αinvolved.␈α Indeed,␈α
the
␈↓ α←␈↓␈↓ β?expert may never know of the models' existence.

␈↓ α←␈↓␈↓ ββ(2)␈↓ β?␈↓¬TEIRESIAS␈↓␈α∃builds␈α⊗and␈α∃updates␈α∃the␈α⊗models␈α∃on␈α∃the␈α⊗basis␈α∃of
␈↓ α←␈↓␈↓ β?␈↓↓experience␈↓,␈α⊃a␈α⊃capability␈α⊃that␈α⊃is␈α⊃intuitively␈α∩appealing.␈α⊃ Every
␈↓ α←␈↓␈↓ β?time␈αa␈αnew␈αrule␈αis␈αadded␈αto␈αthe␈αknowledge␈αbase␈αor␈αan␈αexisting
␈↓ α←␈↓␈↓ β?rule is modified, the appropriate model(s) is recomputed.

␈↓ α←␈↓␈↓ β?There␈α⊃are␈α⊂other␈α⊃implications␈α⊃as␈α⊂well.␈α⊃It␈α⊃was␈α⊂noted␈α⊃above␈α⊃that␈α⊂the
␈↓ α←␈↓model␈αmight␈αbe␈αused␈αto␈αcharacterize␈αmany␈αkinds␈αof␈αfeatures␈αdisplayed␈αby␈αthe
␈↓ α←␈↓rules.␈α∞ The␈α∞correlation␈α∂detection␈α∞operators␈α∞used␈α∞to␈α∂form␈α∞the␈α∞ntuples␈α∂can␈α∞be
␈↓ α←␈↓viewed␈α∞as␈α
demons,␈α∞lying␈α
dormant␈α∞until␈α
enough␈α∞examples␈α∞have␈α
accumulated
␈↓ α←␈↓to␈α∂trigger␈α∂them,␈α∂at␈α∂which␈α∂time␈α∂they␈α∂add␈α∂their␈α∂newfound␈α∂correlation␈α∂to␈α∞the
␈↓ α←␈↓appropriate model.
␈↓ α←␈↓␈↓ β?It␈αwas␈α
also␈αmentioned␈α
above␈αthat␈α
the␈αtrees␈α
around␈αwhich␈α
the␈αmodels
␈↓ α←␈↓are␈α∂organized␈α∞are␈α∂quite␈α∞often␈α∂incomplete,␈α∞since␈α∂there␈α∞may␈α∂be␈α∞too␈α∂few␈α∞rules
␈↓ α←␈↓about␈α∃a␈α∃particular␈α∃attribute.␈α∃ It␈α⊗should␈α∃now␈α∃be␈α∃clear␈α∃that␈α∃this␈α⊗tree␈α∃is
␈↓ α←␈↓constantly␈α⊂growing␈α⊃(and␈α⊂shrinking),␈α⊃exactly␈α⊂paralleling␈α⊃the␈α⊂changes␈α⊃in␈α⊂the
␈↓ α←␈↓knowledge␈αbase.␈α The␈αentire␈αmodel␈αset␈αis␈αthus␈αkept␈αcurrent␈αwith␈αthe␈αgrowing
␈↓ α←␈↓knowledge␈αbase,␈αevolving␈αalong␈αwith␈αit,␈αand␈αreflecting␈αthe␈αshifting␈αpatterns␈αit
␈↓ α←␈↓contains.
␈↓ α←␈↓␈↓ β?The␈αmodels␈αare␈αalso␈αone␈αof␈αthe␈αprimary␈αexamples␈αof␈αa␈α
central␈αtheme
␈↓ α←␈↓of␈αthis␈α
work:␈αhigher␈α
level␈αknowledge.␈α While␈α
the␈αrules␈α
are␈αa␈αrepresentation␈α
of
␈↓ α←␈↓(medical)␈αknowledge,␈αthe␈αmodels␈αare␈αa␈αrepresentation␈αof␈αrules␈αand,␈αhence,␈αare
␈↓ α←␈↓a␈α
representation␈αof␈α
a␈αrepresentation.␈α
 As␈α
generalizations␈αof␈α
a␈αset␈α
of␈αrules,␈α
they
␈↓ α←␈↓may␈α∀suggest␈α∃some␈α∀of␈α∀the␈α∃regularities␈α∀in␈α∀the␈α∃structure␈α∀of␈α∃the␈α∀reasoning
␈↓ α←␈↓embodied in the rules in the knowledge base.
␈↓ α←␈↓␈↓ β?The␈α∩models␈α⊃are␈α∩also␈α⊃``fuzzy.'' ␈α∩Just␈α⊃knowing␈α∩that␈α⊃a␈α∩rule␈α⊃concludes
␈↓ α←␈↓about␈α∞a␈α
particular␈α∞attribute␈α
is␈α∞insufficient␈α
information␈α∞to␈α∞completely␈α
specify
␈↓ α←␈↓its␈α∂premise,␈α⊂but␈α∂it␈α⊂is␈α∂enough␈α∂to␈α⊂permit␈α∂a␈α⊂number␈α∂of␈α⊂plausible␈α∂suggestions.
␈↓ α←␈↓The␈α∀representation␈α∀thus␈α∪requires␈α∀a␈α∀form␈α∪of␈α∀inexactness␈α∀that␈α∀allows␈α∪the
␈↓ α←␈↓specification␈α⊂of␈α∂a␈α⊂range␈α∂of␈α⊂possibilities,␈α∂along␈α⊂with␈α∂the␈α⊂likelihood␈α⊂of␈α∂each.
␈↓ α←␈↓The ␈↓	DESCRIPTION␈↓s of premise and action offer this fuzziness.
␈↓ α←␈↓␈↓ β?There␈αis␈αalso␈α
a␈αcertain␈α``critical␈αmass''␈α
effect.␈α Early␈αin␈αthe␈α
construction
␈↓ α←␈↓of␈α⊃a␈α∩knowledge␈α⊃base,␈α∩there␈α⊃will␈α∩be␈α⊃few␈α∩models␈α⊃and␈α∩their␈α⊃advice␈α∩will␈α⊃be
␈↓ α←␈↓somewhat␈α
unreliable,␈α
since␈αthe␈α
system␈α
is,␈α
in␈αeffect,␈α
generalizing␈α
from␈αtoo␈α
small
␈↓ α←␈↓␈↓5-3␈↓ πbRULE MODEL OVERVIEW    87␈↓

␈↓"β␈↓ α←␈↓a␈α⊃sample.␈α⊃ As␈α∩the␈α⊃knowledge␈α⊃base␈α⊃grows,␈α∩the␈α⊃system's␈α⊃own␈α⊃picture␈α∩of␈α⊃its
␈↓ α←␈↓knowledge␈α⊃gets␈α⊃progressively␈α⊃better,␈α⊃trends␈α⊃become␈α⊃more␈α⊃evident,␈α⊃and␈α⊂the
␈↓ α←␈↓models will make more effective contributions.
␈↓ α←␈↓␈↓ β?The␈α∞way␈α∞in␈α∞which␈α∞the␈α∞models␈α∞are␈α∞used␈α∞implies␈α∞that␈α∞the␈α∞system␈α∞will
␈↓ α←␈↓tend␈αto␈α``hear␈αwhat␈αit␈αexpects␈αto␈αhear,''␈αand␈α``hear␈αwhat␈αit␈αhas␈αheard␈αbefore.'' 
␈↓ α←␈↓This␈α
will␈α
be␈α
explored␈α
in␈α
detail␈α
below,␈α
but␈α
note␈α
that␈α
it␈α
is␈α
not␈α
unlike␈αthe␈α
typical
␈↓ α←␈↓human response in the same situation.
␈↓ α←␈↓␈↓ β?Perhaps␈α
most␈α∞interesting,␈α
the␈α∞presence␈α
of␈α
the␈α∞models␈α
means␈α∞that␈α
the
␈↓ α←␈↓system␈α
has␈α
a␈α
model␈α
of␈α
its␈α
own␈α
knowledge.␈α
 As␈α
will␈α
become␈α
clear␈α
further␈α
on,␈α
in
␈↓ α←␈↓a␈α
primitive␈α
sense␈α
it␈α
``knows␈α
what␈α
it␈α
knows,␈α
and␈α
knows␈α
where␈α
it␈α
is␈α
ignorant''
␈↓ α←␈↓and, in a rudimentary way, can discuss both of these.

␈↓ α←␈↓␈↓α5-3-6    Character and use of the models␈↓
␈↓ α←␈↓␈↓ β?Two␈α
further␈αpoints␈α
remain␈αto␈α
be␈αmade␈α
before␈αreviewing␈α
the␈αdetails␈α
of
␈↓ α←␈↓the␈α⊃trace␈α⊃of␈α⊂system␈α⊃performance.␈α⊃ First,␈α⊂recall␈α⊃that␈α⊃the␈α⊂design␈α⊃and␈α⊃use␈α⊂of
␈↓ α←␈↓models␈αwas␈αdescribed␈αin␈αterms␈αof␈αthree␈αimportant␈αfacets: ␈αThey␈αwere␈αto␈αbe␈αa
␈↓ α←␈↓high-level␈α
description,␈α
they␈α
could␈α
be␈α
used␈α
to␈α
express␈α
expectations␈α
about␈αthe
␈↓ α←␈↓world,␈α
and␈αthey␈α
could␈αbe␈α
used␈αto␈α
guide␈αinterpretation␈α
and␈α``understanding.'' 
␈↓ α←␈↓The␈α
role␈α
of␈α
rule␈αmodels␈α
as␈α
high-level␈α
descriptions␈αshould␈α
be␈α
clear␈α
from␈αthe
␈↓ α←␈↓previous␈α∞discussion.␈α∞ The␈α∞reader␈α∞may␈α
by␈α∞now␈α∞have␈α∞recognized␈α∞their␈α∞use␈α
in
␈↓ α←␈↓expressing␈αexpectations: ␈α
In␈αthe␈α
trace,␈αafter␈α
narrowing␈αdown␈α
the␈αsource␈αof␈α
the
␈↓ α←␈↓problem, ␈↓¬TEIRESIAS␈↓ indicates that

␈↓ α←␈↓∧␈↓ β'I␈α"need␈α"a␈α"rule␈α!to␈α"deduce␈α"that␈α"the␈α"category␈α!of
␈↓ α←␈↓∧␈↓ β'ORGANISM-1 is enterobacteriaceae.

␈↓ α←␈↓This␈α
is␈α
both␈α
a␈α
set␈α
of␈α
instructions␈α
to␈α
the␈α
user␈α
about␈α
how␈α
to␈α
fix␈α
the␈α
problem
␈↓ α←␈↓and␈αan␈α
indication␈αof␈α
the␈αsort␈α
of␈αrule␈α
the␈αsystem␈α
will␈αexpect--it␈α
indicates␈αthe
␈↓ α←␈↓rule model that will be employed.
␈↓ α←␈↓␈↓ β?The␈αuse␈αof␈αthe␈αmodels␈αas␈αa␈αguide␈αto␈αunderstanding␈αis␈αthe␈αfinal␈αpoint.
␈↓ α←␈↓With␈α
the␈α
expectations␈α∞that␈α
arise␈α
from␈α
doing␈α∞acquisition␈α
in␈α
the␈α
context␈α∞of␈α
a
␈↓ α←␈↓shortcoming␈α∪in␈α∪the␈α∪knowledge␈α∪base,␈α∪the␈α∪system␈α∪is␈α∪not␈α∪entering␈α∀the␈α∪rule
␈↓ α←␈↓interpretation␈α
process␈α
blindly.␈α
 In␈α
terms␈α
of␈α
the␈α
three␈αmethodologies␈α
mentioned
␈↓ α←␈↓earlier,␈α⊃then,␈α⊃the␈α⊂␈↓↓descriptive␈↓␈α⊃approach␈α⊃is␈α⊂inappropriate.␈α⊃ But␈α⊃the␈α⊂program
␈↓ α←␈↓should␈α
not␈α∞be␈α
restricted␈α
to␈α∞understanding␈α
only␈α
rules␈α∞that␈α
look␈α∞like␈α
previous
␈↓ α←␈↓rules␈α∂in␈α∞the␈α∂model,␈α∂since␈α∞the␈α∂new␈α∞ones␈α∂may␈α∂be␈α∞different␈α∂from␈α∂all␈α∞previous
␈↓ α←␈↓examples.␈α∪ It␈α∪would␈α∪thus␈α∪be␈α∪too␈α∪inflexible␈α∪to␈α∪use␈α∪a␈α∩␈↓↓verification␈↓-oriented
␈↓ α←␈↓approach,␈αand␈αthe␈α␈↓↓recognition␈↓␈αtechnique␈αappears␈αto␈αbe␈αjust␈αright.␈α There␈αare
␈↓ α←␈↓expectations,␈α∞but␈α
it␈α∞is␈α∞not␈α
certain␈α∞that␈α
they␈α∞will␈α∞be␈α
fulfilled.␈α∞Hence␈α∞the␈α
rule
␈↓ α←␈↓text␈α∂is␈α∞processed␈α∂``bottom␈α∞up,''␈α∂but␈α∞with␈α∂a␈α∞strong␈α∂set␈α∞of␈α∂biases␈α∂about␈α∞where
␈↓ α←␈↓that␈αprocessing␈αshould␈αeventually␈αlead.␈α In␈αaddition,␈αthe␈αrule␈αmodels,␈αlike␈αthe
␈↓ α←␈↓polyhedra␈α
models,␈α
will␈α
be␈α
used␈α
to␈α
direct␈α
the␈α
processing␈α
of␈α
the␈α
text␈α
and␈α
to␈α
help
␈↓ α←␈↓identify both noise and gaps in the signal.
␈↓ α←␈↓␈↓88    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓␈↓α5-4    HOW IT ALL WORKS␈↓
␈↓ α←␈↓␈↓ β?At␈αthis␈αpoint␈αwe␈αbegin␈αthe␈αtrace␈αagain,␈αreviewing␈αit␈αpiece␈αby␈αpiece,␈αas
␈↓ α←␈↓a␈α
background␈α∞for␈α
discussing␈α
some␈α∞of␈α
the␈α
more␈α∞interesting␈α
ideas␈α∞in␈α
␈↓¬TEIRESIAS␈↓
␈↓ α←␈↓and␈αas␈α
a␈αway␈α
of␈αexploring␈α
features␈αnot␈α
yet␈αillustrated.␈α
 To␈αavoid␈αthe␈α
necessity
␈↓ α←␈↓of␈α∂flipping␈α∂pages,␈α∂the␈α∂relevant␈α∂sections␈α∂of␈α∂the␈α∂trace␈α∂have␈α⊂been␈α∂reproduced
␈↓ α←␈↓below, set off by horizontal lines.
␈↓ α←␈↓␈↓ β?It␈α∩is␈α∩both␈α⊃difficult␈α∩and␈α∩not␈α⊃particularly␈α∩informative␈α∩to␈α∩attempt␈α⊃to
␈↓ α←␈↓review␈α∞all␈α∞the␈α∞options␈α∞in␈α∞the␈α∞system,␈α∞so␈α∞only␈α∞the␈α∞more␈α∂interesting␈α∞examples
␈↓ α←␈↓will␈αbe␈α
explored.␈α In␈αgeneral,␈α
however,␈αthe␈αproper␈α
facilities␈αexist␈αfor␈α
handling
␈↓ α←␈↓all␈α⊂cases,␈α⊂and␈α⊂short␈α⊂of␈α⊂purposeful␈α⊂attempts␈α⊂to␈α⊂supply␈α⊃inconsistent␈α⊂answers,
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓ takes the appropriate action.
␈↓ α←␈↓␈↓ β?For␈α⊂the␈α⊂sake␈α⊃of␈α⊂consistency,␈α⊂we␈α⊂will␈α⊃work␈α⊂with␈α⊂the␈α⊃single␈α⊂example
␈↓ α←␈↓shown␈αin␈αthe␈α
original␈αtrace.␈α It␈αis␈α
occasionally␈αnecessary,␈αhowever,␈αto␈α
illustrate
␈↓ α←␈↓additional␈αpoints␈αby␈αexamining␈αthe␈αsystem's␈αaction␈αon␈αhypothetical␈αresponses
␈↓ α←␈↓from the user.  These will be clearly indicated.
␈↓ α←␈↓␈↓ β?There are seven general phases to the rule acquisition process:

␈↓ α←␈↓␈↓ ββ(a)␈↓ β?tracking down the bug,
␈↓ α←␈↓␈↓ ββ(b)␈↓ β?deciphering the English text of the new rule,
␈↓ α←␈↓␈↓ ββ(c)␈↓ β?checking preliminary results,
␈↓ α←␈↓␈↓ ββ(d)␈↓ β?``second guessing,''
␈↓ α←␈↓␈↓ ββ(e)␈↓ β?final checkout,
␈↓ α←␈↓␈↓ ββ(f)␈↓ β?bookkeeping, and
␈↓ α←␈↓␈↓ ββ(g)␈↓ β?rerunning the consultation.

␈↓ α←␈↓Each is reviewed in turn.

␈↓ α←␈↓␈↓α5-4-1    Tracking down the bug␈↓
␈↓ α←␈↓␈↓ ¬GI␈αwill␈αnot␈αbe␈αpersuaded␈αto␈αleave␈αoff␈αthe␈α
chance␈αof
␈↓ α←␈↓␈↓ ¬Gfinding out the whole thing clearly.
␈↓"β␈↓ α←␈↓␈↓ λ␈lines 1065-1066
␈↓ α←␈↓␈↓ β?The␈αperformance␈αprogram's␈αdiagnosis␈αis␈α
used␈αas␈αan␈αentry␈αpoint␈αto␈α
the
␈↓ α←␈↓debugging␈α∩routine.␈α∩ This␈α∩is␈α∩a␈α∩logical␈α∩evaluation␈α∩point␈α∩since␈α∩most␈α∩of␈α∩the
␈↓ α←␈↓consultation␈αis␈αaimed␈αtoward␈αdetermining␈αit.␈α It␈αis␈αpossible␈αthat␈αthe␈αdiagnosis
␈↓ α←␈↓may␈α⊃be␈α⊃correct␈α⊃due␈α⊃to␈α⊃offsetting␈α⊂errors␈α⊃in␈α⊃the␈α⊃course␈α⊃of␈α⊃the␈α⊂consultation.
␈↓ α←␈↓However,␈α∞since␈α∞the␈α∞expert␈α∞can␈α
interrupt␈α∞the␈α∞consultation␈α∞at␈α∞anytime␈α
during
␈↓ α←␈↓the␈α
processing␈αto␈α
use␈α
the␈αexplanation␈α
and␈α
acquisition␈αroutines,␈α
we␈α
make␈αthe
␈↓ α←␈↓plausible␈α(but␈αimperfect)␈αassumption␈αthat␈αif␈αhe␈αgot␈αthis␈αfar␈αand␈αthe␈αdiagnosis
␈↓ α←␈↓is correct, then everything else is correct also.
␈↓ α←␈↓␈↓5-4␈↓ λ HOW IT ALL WORKS    89␈↓

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬My therapy recommendation will be based on the following probable infection(s) and
␈↓"β␈↓ α←␈↓¬potential causative organism(s):

␈↓"β␈↓ α←␈↓¬INFECTION-1 is PRIMARY-BACTEREMIA
␈↓"β␈↓ α←␈↓¬    <Item 1>   E.COLI [ORGANISM-1]
␈↓"β␈↓ α←␈↓¬    <Item 2>  PSEUDOMONAS-AERUGINOSA [ORGANISM-1]

␈↓"β␈↓ α←␈↓¬Are all these identities correct?
␈↓"β␈↓ α←␈↓¬++**␈↓π N␈↓¬
␈↓"β␈↓ α←␈↓¬Are there any identities missing?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬
␈↓"β␈↓ α←␈↓¬Which one(s)?
␈↓"β␈↓ α←␈↓¬++**␈↓π KLEBSIELLA-PNEUMONIAE   PROTEUS-NON-MIRABILIS␈↓¬
␈↓"β␈↓ α←␈↓¬Are there any identities that shouldn't have appeared?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬
␈↓"β␈↓ α←␈↓¬Which one(s)?
␈↓"β␈↓ α←␈↓¬++**␈↓π PSEUDOMONAS-AERUGINOSA
␈↓"β␈↓ α←␈↓π_______________________________________

␈↓ α←␈↓␈↓ β?The␈αfocusing␈αprocess␈αbegins␈αimmediately,␈αas␈αthe␈αexpert␈αis␈αrequired␈αto
␈↓ α←␈↓specify␈α
exactly␈α
why␈α
he␈α
doesn't␈α
like␈α
the␈α
diagnosis␈α
shown.␈α
 There␈α∞must␈α
either
␈↓ α←␈↓be␈αan␈αinappropriate␈αanswer␈αpresent;␈αone␈αthat␈αis␈αmissing;␈αor,␈αas␈αin␈αthe␈αcurrent
␈↓ α←␈↓example, both.

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬Do you want to
␈↓"β␈↓ α←␈↓¬      1] track down the problem yourself, or
␈↓"β␈↓ α←␈↓¬      2] allow me to guide the debugging, or
␈↓"β␈↓ α←␈↓¬      3] ignore it for now and fix it another time?
␈↓"β␈↓ α←␈↓¬Type 1, 2 or 3
␈↓"β␈↓ α←␈↓¬++**␈↓π 2
␈↓"β␈↓ α←␈↓π_______________________________________

␈↓ α←␈↓␈↓ β?Alternative␈α∂#1␈α∂was␈α∞explored␈α∂in␈α∂chapter␈α∂3,␈α∞and␈α∂#2␈α∂is␈α∂reviewed␈α∞here.
␈↓ α←␈↓Alternative␈α⊂#3␈α⊃is␈α⊂part␈α⊂of␈α⊃the␈α⊂emphasis␈α⊂on␈α⊃human␈α⊂engineering.␈α⊃ Since␈α⊂the
␈↓ α←␈↓process␈αof␈αtracking␈αdown␈αand␈αfixing␈αthe␈αbug␈αis␈αa␈αsubstantive␈αtask,␈αthe␈α
expert
␈↓ α←␈↓is␈α
given␈αa␈α
chance␈αto␈α
avoid␈α
it.␈α If␈α
he␈αtakes␈α
this␈α
option,␈αthe␈α
system␈αreminds␈α
him
␈↓ α←␈↓to␈α``try␈α
to␈αremember␈αto␈α
fix␈αit␈αsome␈α
time''␈αand␈α(quietly)␈α
composes␈αand␈α
sends␈αa
␈↓ α←␈↓note␈αto␈αa␈αbug-reporting␈αfile,␈αgiving␈αthe␈αcase␈αnumber␈αand␈αthe␈αlist␈αof␈αincorrect
␈↓ α←␈↓or␈αmissing␈αanswers␈αindicated␈αby␈αthe␈αuser.␈α This␈αfile␈αis␈αroutinely␈αexamined␈αby
␈↓ α←␈↓system␈α∂programmers,␈α∂who␈α∂can␈α∂either␈α∂fix␈α∂the␈α∂problem␈α∂or␈α∂simply␈α∂make␈α∞sure
␈↓ α←␈↓that it gets attention later.

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬We'll start with the PSEUDOMONAS-AERUGINOSA first.

␈↓"β␈↓ α←␈↓¬The following rules were used in deducing that the identity of ORGANISM-1 is
␈↓"β␈↓ α←␈↓¬pseudomonas-aeruginosa

␈↓"β␈↓ α←␈↓¬RULE184
␈↓"β␈↓ α←␈↓¬-------
␈↓"β␈↓ α←␈↓¬Since [1.1] the category of ORGANISM-1 is not known
␈↓"β␈↓ α←␈↓¬      [1.2] the gram stain of ORGANISM-1 is gramneg
␈↓"β␈↓ α←␈↓¬      [1.3] the morphology of ORGANISM-1 is rod
␈↓"β␈↓ α←␈↓¬      [1.4] the aerobicity of ORGANISM-1 is facultative
␈↓"β␈↓ α←␈↓¬ There is weakly suggestive evidence (.3) that the identity of ORGANISM-1 is
␈↓ α←␈↓␈↓90    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓¬ pseudomonas-aeruginosa

␈↓"β␈↓ α←␈↓¬Is this rule correct?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬
␈↓"β␈↓ α←␈↓¬Should its premise have failed for this case?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬
␈↓"β␈↓ α←␈↓¬Which clause of the premise should have been false? [give #]
␈↓"β␈↓ α←␈↓¬++**␈↓π 1
␈↓"β␈↓ α←␈↓π_______________________________________

␈↓ α←␈↓␈↓ β?The␈α_last␈α_three␈α_questions␈α_directed␈α_to␈α_the␈α_user␈α→demonstrate␈α_the
␈↓ α←␈↓methodical␈α∂unwinding␈α∂process␈α⊂that␈α∂forces␈α∂criticism␈α⊂to␈α∂be␈α∂very␈α⊂specific.␈α∂ At
␈↓ α←␈↓each␈α⊂point,␈α⊂the␈α⊃expert␈α⊂must␈α⊂either␈α⊂approve␈α⊃of␈α⊂the␈α⊂rules␈α⊂invoked␈α⊃and␈α⊂the
␈↓ α←␈↓values␈α
obtained␈α
or␈α
indicate␈α
which␈αone␈α
was␈α
in␈α
error␈α
and␈α
supply␈αthe␈α
correction.
␈↓ α←␈↓An␈α
analysis␈α
of␈α
the␈α
performance␈α
program's␈α
control␈α
structure␈α
indicates␈α
that␈α
if
␈↓ α←␈↓the␈α
value␈α
of␈α
an␈α
attribute␈α
A␈α
is␈α
incorrect,␈α
then␈α
(if␈α
S␈α
is␈α
the␈α
set␈α
of␈α
rules␈α
in␈αthe
␈↓ α←␈↓knowledge base that conclude about A) either:

␈↓ α←␈↓␈↓ ββ(a)␈↓ β?Some␈α∀rule␈α∀in␈α∪S␈α∀which␈α∀was␈α∀executed␈α∪had␈α∀an␈α∀error␈α∀in␈α∪its
␈↓ α←␈↓␈↓ β?conclusion, or

␈↓ α←␈↓␈↓ ββ(b)␈↓ β?some␈αrule␈αin␈αS␈αwas␈αnot␈αexecuted␈αand␈αshould␈αhave␈αbeen␈α
(either
␈↓ α←␈↓␈↓ β?because␈α
of␈α∞an␈α
error␈α
in␈α∞its␈α
premise␈α
or␈α∞because␈α
a␈α
clause␈α∞of␈α
the
␈↓ α←␈↓␈↓ β?premise incorrectly evaluated to ``false''), or

␈↓ α←␈↓␈↓ ββ(c)␈↓ β?some␈αrule␈αin␈αS␈αwas␈αexecuted␈αand␈αshould␈αnot␈αhave␈αbeen␈α
(either
␈↓ α←␈↓␈↓ β?because␈α
of␈α∞an␈α
error␈α
in␈α∞its␈α
premise␈α
or␈α∞because␈α
a␈α
clause␈α∞of␈α
the
␈↓ α←␈↓␈↓ β?premise incorrectly evaluated to ``true''), or

␈↓ α←␈↓␈↓ ββ(d)␈↓ β?the set S is incomplete, and one or more rules is missing.


␈↓ α←␈↓The␈α∞routines␈α∞in␈α∞␈↓¬TEIRESIAS␈↓␈α∞reflect␈α∞this␈α∞analysis␈α∞and␈α∞can␈α∞account␈α∞for␈α∂all␈α∞error
␈↓ α←␈↓sources␈αwithin␈αthe␈αcurrent␈α
control␈αstructure.␈α Thus,␈αhad␈αthe␈α
expert␈αindicated
␈↓ α←␈↓that␈α
rule␈α184␈α
was␈α
incorrect,␈αhe␈α
would␈α
have␈αbeen␈α
invited␈α
to␈αuse␈α
the␈αrule␈α
editor
␈↓ α←␈↓to␈α∃fix␈α⊗it.␈α∃ Had␈α⊗he␈α∃indicated␈α⊗that␈α∃it␈α∃was␈α⊗correct␈α∃and␈α⊗that␈α∃it␈α⊗was␈α∃not
␈↓ α←␈↓inappropriately␈α∞invoked,␈α
the␈α∞assumption␈α
would␈α∞have␈α
been␈α∞that␈α
there␈α∞was␈α
a
␈↓ α←␈↓rule␈α
missing␈α∞that␈α
concluded␈α
negatively␈α∞about␈α
the␈α
presence␈α∞of␈α
pseudomonas-
␈↓ α←␈↓aeruginosa␈α∪(the␈α∩missing␈α∪rule␈α∩would␈α∪offset␈α∩the␈α∪action␈α∩of␈α∪184␈α∪and,␈α∩hence,
␈↓ α←␈↓remove the inappropriate identity).
␈↓ α←␈↓␈↓ β?The␈α
possibility␈α
remains,␈α
however,␈α
that␈α
the␈α
framework␈α
itself␈α
may␈α
break
␈↓ α←␈↓down␈α∀some␈α∀day.␈α∀ That␈α∪is,␈α∀the␈α∀expert␈α∀may␈α∪claim␈α∀that␈α∀all␈α∀the␈α∀rules␈α∪are
␈↓ α←␈↓individually␈α⊃correct,␈α⊃that␈α⊃they␈α⊃were␈α⊃all␈α⊃executed␈α⊃correctly,␈α⊃that␈α⊃no␈α∩rule␈α⊃is
␈↓ α←␈↓missing,␈α
and␈α
yet␈αthe␈α
final␈α
answer␈αis␈α
still␈α
incorrect.␈α
 This␈αis␈α
most␈α
likely␈αto␈α
come
␈↓ α←␈↓about␈α⊃as␈α⊃a␈α⊂result␈α⊃of␈α⊃the␈α⊂possible␈α⊃inapplicability␈α⊃in␈α⊂a␈α⊃new␈α⊃domain␈α⊃of␈α⊂the
␈↓ α←␈↓performance␈α⊂program's␈α⊂control␈α⊂structure␈α⊂or␈α⊂certainty␈α⊂factor␈α⊂mechanism.␈α⊂ If
␈↓ α←␈↓the␈α∂expert␈α⊂does␈α∂indicate␈α⊂that␈α∂none␈α⊂of␈α∂the␈α⊂established␈α∂classifications␈α⊂of␈α∂the
␈↓ α←␈↓␈↓5-4␈↓ λ HOW IT ALL WORKS    91␈↓

␈↓"β␈↓ α←␈↓problem␈αis␈αcorrect,␈α␈↓¬TEIRESIAS␈↓␈αwill␈αindicate␈αits␈αinability␈αto␈αhelp␈αfurther␈αand␈αuse
␈↓ α←␈↓the bug reporting mechanism noted above to make a note of the problem.
␈↓ α←␈↓␈↓ β?In␈αthis␈αcase,␈αhowever,␈αthe␈αexpert␈αhas␈αindicated␈αthat␈αthe␈αfirst␈αclause␈αof
␈↓ α←␈↓rule␈α∀184␈α∪(``␈↓	the␈α∀category␈α∀of␈α∪the␈α∀organism␈α∪is␈α∀not␈α∀known␈↓'')␈α∪incorrectly
␈↓ α←␈↓evaluated to ``true.''  The process continues from there.

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬No conclusions were reached concerning the category of ORGANISM-1.

␈↓"β␈↓ α←␈↓¬Should it have been possible to conclude a category?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬

␈↓"β␈↓ α←␈↓¬Which one?
␈↓"β␈↓ α←␈↓¬++**␈↓π ENTEROBACTERIACEAE␈↓¬

␈↓"β␈↓ α←␈↓¬There were no rules successfully invoked in the attempt to deduce that the category
␈↓"β␈↓ α←␈↓¬of ORGANISM-1 is enterobacteriaceae.

␈↓"β␈↓ α←␈↓¬The following (potentially applicable) rules were NOT executed because the
␈↓"β␈↓ α←␈↓¬starred [**] clause failed:

␈↓"β␈↓ α←␈↓¬RULE116
␈↓"β␈↓ α←␈↓¬-------
␈↓"β␈↓ α←␈↓¬If    1) the identity of ORGANISM-1 is not known
␈↓"β␈↓ α←␈↓¬   ** 2) the gram stain of ORGANISM-1 is not known **
␈↓"β␈↓ α←␈↓¬      3) the morphology of ORGANISM-1 is not known
␈↓"β␈↓ α←␈↓¬      4) the site of CULTURE-1 is csf
␈↓"β␈↓ α←␈↓¬      5) the infection is meningitis
␈↓"β␈↓ α←␈↓¬      6) the age (in years) of Albert Hall is less than or equal to .17
␈↓"β␈↓ α←␈↓¬then  There is weakly suggestive evidence (.33) that the category of ORGANISM-1 is
␈↓"β␈↓ α←␈↓¬enterobacteriaceae

␈↓"β␈↓ α←␈↓¬RULE050
␈↓"β␈↓ α←␈↓¬-------
␈↓"β␈↓ α←␈↓¬If    1) the morphology of ORGANISM-1 is rod
␈↓"β␈↓ α←␈↓¬      2) the gram stain of ORGANISM-1 is gramneg
␈↓"β␈↓ α←␈↓¬      3) the aerobicity of ORGANISM-1 is facultative
␈↓"β␈↓ α←␈↓¬   ** 4) the infection with ORGANISM-1 was acquired while the patient was
␈↓"β␈↓ α←␈↓¬         hospitalized **
␈↓"β␈↓ α←␈↓¬then  There is suggestive evidence (.7) that the category of ORGANISM-1 is
␈↓"β␈↓ α←␈↓¬      enterobacteriaceae
␈↓"β␈↓ α←␈↓¬_______________________________________

␈↓ α←␈↓␈↓ β?Since␈α∞the␈α∞need␈α∞to␈α∞deal␈α∞with␈α∞large␈α∞numbers␈α∞of␈α∞rules␈α∞was␈α∞anticipated,
␈↓ α←␈↓the␈α∃rule␈α∃unwinding␈α∃process␈α∃is␈α∃kept␈α∃as␈α∃constrained␈α∃as␈α⊗possible.␈α∃ Before
␈↓ α←␈↓displaying␈αall␈αthe␈αrules␈αthat␈αmight␈αhave␈αconcluded␈αany␈αvalue␈αof␈αcategory,␈αthe
␈↓ α←␈↓expert␈α∞is␈α∞asked␈α∞which␈α∞category␈α∞it␈α
should␈α∞have␈α∞been.␈α∞ Then,␈α∞only␈α∞rules␈α
that
␈↓ α←␈↓are appropriate to his answer need be considered.
␈↓ α←␈↓␈↓92    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬Are all these rules correct?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬
␈↓"β␈↓ α←␈↓¬Should any of them have been successfully invoked?
␈↓"β␈↓ α←␈↓¬+**␈↓π N␈↓¬
␈↓"β␈↓ α←␈↓¬Is there a rule missing?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬

␈↓"β␈↓ α←␈↓¬Ok, then I need a rule that would allow me to deduce that the category of ORGANISM-1
␈↓"β␈↓ α←␈↓¬is enterobacteriaceae.

␈↓"β␈↓ α←␈↓¬Would you like to give me such a rule now?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬
␈↓"β␈↓ α←␈↓¬Good...

␈↓"β␈↓ α←␈↓¬The new rule will be called RULE383
␈↓"β␈↓ α←␈↓¬ If:    1 -  ␈↓πTHE PATIENT'S INFECTION IS PRIMARY-BACTEREMIA␈↓¬
␈↓"β␈↓ α←␈↓¬   and  2 -  ␈↓πTHE SITE IS ONE OF THE STERILE SITES␈↓¬
␈↓"β␈↓ α←␈↓¬   and  3 -
␈↓"β␈↓ α←␈↓¬ Then:  1 -  ␈↓πTHERE IS EVIDENCE (.8) THAT THE CATEGORY IS ENTEROBACTERIACEAE␈↓¬
␈↓"β␈↓ α←␈↓¬   and  2 -

␈↓"β␈↓ α←␈↓¬This may take a few moments...
␈↓"β␈↓ α←␈↓¬_______________________________________

␈↓ α←␈↓␈↓ β?The␈α∞problem␈α∞has␈α∞finally␈α∞been␈α∞tracked␈α∞down␈α∞and␈α∞characterized␈α∞with
␈↓ α←␈↓an␈α∞indication␈α∞of␈α∞exactly␈α∞what␈α∞sort␈α
of␈α∞rule␈α∞is␈α∞needed.␈α∞ As␈α∞noted␈α∞earlier,␈α
this
␈↓ α←␈↓characterization␈αalso␈αserves␈αthe␈α
purpose␈αof␈αsetting␈αup␈α
␈↓¬TEIRESIAS␈↓'s␈αexpectations
␈↓ α←␈↓about␈α
the␈αsort␈α
of␈αrule␈α
it␈αis␈α
about␈αto␈α
receive.␈α Since␈α
there␈αis␈α
not␈α
necessarily␈αa
␈↓ α←␈↓rule␈α∩model␈α∩for␈α∩every␈α∩characterization,␈α⊃the␈α∩system␈α∩chooses␈α∩the␈α∩model␈α⊃that
␈↓ α←␈↓matches␈α⊃most␈α⊃closely.␈α⊃ This␈α⊂is␈α⊃done␈α⊃by␈α⊃starting␈α⊃at␈α⊂the␈α⊃top␈α⊃of␈α⊃the␈α⊃tree␈α⊂of
␈↓ α←␈↓models␈α
and␈α
descending␈α
until␈α
either␈α
reaching␈α
a␈α
model␈α
of␈α
the␈α
desired␈α
type␈αor
␈↓ α←␈↓encountering␈α⊂a␈α⊃leaf␈α⊂of␈α⊂the␈α⊃tree.␈α⊂ In␈α⊃this␈α⊂case,␈α⊂the␈α⊃process␈α⊂descends␈α⊃to␈α⊂the
␈↓ α←␈↓second␈α
level␈α
(the␈α
␈↓	CATEGORY-IS␈↓␈α
model),␈α
notices␈α
that␈α
there␈α
is␈αno␈α
␈↓	CATEGORY-IS-
␈↓ α←␈↓	ENTEROBACTERIACEAE␈↓␈α⊂model␈α⊃at␈α⊂the␈α⊃next␈α⊂level,␈α⊂and␈α⊃settles␈α⊂for␈α⊃the␈α⊂former.
␈↓ α←␈↓This␈α
technique␈α
is␈α∞used␈α
in␈α
several␈α∞places␈α
throughout␈α
the␈α∞knowledge␈α
transfer
␈↓ α←␈↓process␈α⊃and,␈α⊂in␈α⊃general,␈α⊂supplies␈α⊃the␈α⊂model␈α⊃that␈α⊂best␈α⊃matches␈α⊃the␈α⊂current
␈↓ α←␈↓requirements.␈α
 Note␈α
that␈α
it␈α
can␈α
deal␈αwith␈α
varying␈α
levels␈α
of␈α
specificity␈α
in␈αthe
␈↓ α←␈↓stated␈α∪expectations.␈α∩ If,␈α∪for␈α∪instance,␈α∩the␈α∪system␈α∪had␈α∩known␈α∪only␈α∪that␈α∩it
␈↓ α←␈↓expected␈α∞a␈α∞rule␈α∂that␈α∞concluded␈α∞affirmatively␈α∂about␈α∞category,␈α∞it␈α∂would␈α∞have
␈↓ α←␈↓descended just that far in the model tree and looked no further.

␈↓ α←␈↓␈↓α5-4-2    Deciphering the English text␈↓
␈↓ α←␈↓␈↓ β?It␈αwas␈αsuggested␈αearlier␈αthat␈αinterpreting␈αthe␈αnatural␈αlanguage␈αtext␈αof
␈↓ α←␈↓the␈α⊃rule␈α⊃can␈α⊂be␈α⊃viewed␈α⊃as␈α⊂a␈α⊃``recognition''␈α⊃process␈α⊂in␈α⊃which␈α⊃the␈α⊃data␈α⊂are
␈↓ α←␈↓allowed␈α⊃to␈α⊂suggest␈α⊃interpretations,␈α⊂but␈α⊃the␈α⊂system␈α⊃maintains␈α⊃certain␈α⊂biases
␈↓ α←␈↓about␈α⊗which␈α⊗interpretation␈α⊗is␈α⊗likely␈α⊗to␈α⊗be␈α⊗correct.␈α⊗ ␈↓¬TEIRESIAS␈↓␈α↔does␈α⊗this,
␈↓ α←␈↓generating␈αall␈αconsistent␈αinterpretations␈αof␈αeach␈αline␈αof␈αEnglish␈αtext␈αand␈αthen
␈↓ α←␈↓evaluating␈α⊂each␈α⊂interpretation␈α⊂in␈α⊂the␈α⊂light␈α⊂of␈α⊂the␈α⊂biases␈α⊂expressed␈α⊃by␈α⊂the
␈↓ α←␈↓choice␈α∂of␈α⊂a␈α∂specific␈α⊂rule␈α∂model.␈α⊂ The␈α∂interpretations␈α⊂suggested␈α∂by␈α⊂the␈α∂text
␈↓ α←␈↓(data-driven,␈α∂``bottom␈α∞up''␈α∂mode)␈α∞are␈α∂thus␈α∞intersected␈α∂with␈α∂the␈α∞expectations
␈↓ α←␈↓(hypothesis-driven, ``top down'' mode) provided by the debugging process.
␈↓ α←␈↓␈↓5-4␈↓ λ HOW IT ALL WORKS    93␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃interpretation␈α⊃process␈α⊃works␈α⊃in␈α⊃a␈α⊃strictly␈α⊃line-by-line␈α⊃fashion,
␈↓ α←␈↓processing␈αeach␈αline␈αof␈αtext␈αindependently.␈α This␈αmethod␈αis␈αa␈αsource␈αof␈αsome
␈↓ α←␈↓deficiencies,␈α∂some␈α⊂of␈α∂which␈α⊂are␈α∂trivially␈α∂fixed,␈α⊂while␈α∂others␈α⊂are␈α∂superficial
␈↓ α←␈↓manifestations␈α⊗of␈α⊗interesting␈α⊗and␈α↔complex␈α⊗problems.␈α⊗ Each␈α⊗of␈α↔them␈α⊗is
␈↓ α←␈↓discussed in subsequent sections below.
␈↓ α←␈↓␈↓ β?Deciphering the text occurs in four stages:

␈↓ α←␈↓␈↓ ββ(a)␈↓ β?preprocessing the text,
␈↓ α←␈↓␈↓ ββ(b)␈↓ β?checking the rule model,
␈↓ α←␈↓␈↓ ββ(c)␈↓ β?generating the set of plausible ␈↓¬LISP␈↓ interpretations, and
␈↓ α←␈↓␈↓ ββ(d)␈↓ β?scoring the interpretations by reference to the rule model.

␈↓ α←␈↓␈↓ β?As␈α
will␈α
become␈α
clear,␈α
our␈α
approach␈α
to␈α
natural␈α
language␈α
is␈α
very␈α
simple,
␈↓ α←␈↓yet␈α
powerful␈α∞enough␈α
to␈α∞support␈α
the␈α
performance␈α∞required.␈α
 The␈α∞problem␈α
is
␈↓ α←␈↓made␈α
easier,␈α
of␈α
course,␈α
by␈α
the␈α
fact␈α
that␈α
we␈α
are␈α
dealing␈α
with␈α
a␈α
small␈α
amount␈α
of
␈↓ α←␈↓text␈α∩written␈α∩in␈α∩a␈α∩semi-formal␈α∩technical␈α∩language,␈α∩rather␈α∩than␈α∪with␈α∩large
␈↓ α←␈↓amounts␈α≤of␈α≤text␈α≥in␈α≤unrestricted␈α≤dialog.␈α≥ Even␈α≤so,␈α≤the␈α≥problem␈α≤of
␈↓ α←␈↓interpretation␈α∨is␈α∨substantial.␈α∨ The␈α≡source␈α∨of␈α∨␈↓¬TEIRESIAS␈↓'s␈α∨power␈α≡and
␈↓ α←␈↓performance␈αlies␈αin␈αits␈αuse␈αof␈αa␈αmultiplicity␈αof␈αknowledge␈αsources.␈α These␈αare
␈↓ α←␈↓listed␈αand␈αdescribed␈αbriefly␈αbelow;␈αtheir␈αuse␈αis␈αexplored␈αin␈αmore␈αdetail␈αin␈αthe
␈↓ α←␈↓sections␈αthat␈αfollow.␈α Since␈αmuch␈αof␈αthe␈αinterpretation␈αprocess␈αcan␈αbe␈αviewed
␈↓ α←␈↓in␈αterms␈αof␈αforming␈αhypotheses␈α(interpretations)␈αfrom␈αa␈αset␈αof␈αdata␈α(the␈αtext),
␈↓ α←␈↓each knowledge source is also labeled in these terms.

␈↓ α←␈↓␈↓ ββ␈↓ β?    Data-driven knowledge sources:

␈↓ α←␈↓␈↓ ββ(1)␈↓ β?␈↓↓Connotations␈α∃of␈α∃individual␈α∃words␈↓␈α∃(data␈α∃interpretation).␈α∃ As
␈↓ α←␈↓␈↓ β?explained␈αin␈α
more␈αdetail␈α
below,␈αeach␈α
English␈αword␈α
may␈αhave
␈↓ α←␈↓␈↓ β?associated␈α↔with␈α_it␈α↔a␈α↔number␈α_of␈α↔connotations␈α_of␈α↔varying
␈↓ α←␈↓␈↓ β?strength.␈α≡ These␈α∨indicate␈α≡attributes,␈α≡objects,␈α∨values,␈α≡or
␈↓ α←␈↓␈↓ β?predicate␈α⊗functions␈α⊗to␈α⊗which␈α⊗the␈α⊗word␈α⊗may␈α⊗plausibly␈α∃be
␈↓ α←␈↓␈↓ β?referring.

␈↓ α←␈↓␈↓ ββ(2)␈↓ β?␈↓↓Degree␈α∪of␈α∪ambiguity␈α∪of␈α∪individual␈α∪words␈↓␈α∪(ambiguity␈α∪of␈α∩the
␈↓ α←␈↓␈↓ β?data).␈α≥ This␈α≡is␈α≥used␈α≥to␈α≡constrain␈α≥the␈α≥search␈α≡for␈α≥an
␈↓ α←␈↓␈↓ β?interpretation of the text.

␈↓ α←␈↓␈↓ ββ(3)␈↓ β?␈↓↓Function␈α∞template␈↓␈α∞(structure␈α∂of␈α∞the␈α∞hypothesis).␈α∞ As␈α∂noted␈α∞in
␈↓ α←␈↓␈↓ β?chapter␈α∂2,␈α∂there␈α∂is␈α∞associated␈α∂with␈α∂each␈α∂predicate␈α∂function␈α∞a
␈↓ α←␈↓␈↓ β?template␈α≤indicating␈α≤the␈α≤order␈α≠and␈α≤generic␈α≤type␈α≤of␈α≠its
␈↓ α←␈↓␈↓ β?arguments.␈α∞ Generating␈α
code␈α∞is␈α∞essentially␈α
a␈α∞process␈α∞of␈α
filling
␈↓ α←␈↓␈↓ β?in␈αthis␈αtemplate;␈αit␈αthus␈α
provides␈αa␈αprimary␈αsource␈αof␈α
direction
␈↓ α←␈↓␈↓ β?for the interpretation process.

␈↓ α←␈↓␈↓ ββ(4)␈↓ β?␈↓↓Degree␈αof␈αsuccess␈αof␈αthe␈αtemplate␈αcompletion␈αprocess␈↓␈α
(degree␈αof
␈↓ α←␈↓␈↓94    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓␈↓ β?success␈α∞of␈α∞hypothesis␈α∞construction).␈α∞ The␈α∞expert␈α∞may␈α∞be␈α∞terse
␈↓ α←␈↓␈↓ β?enough␈α
in␈αhis␈α
rule␈α
statement␈αthat␈α
the␈αcontents␈α
of␈α
certain␈αslots
␈↓ α←␈↓␈↓ β?in␈α⊂the␈α⊂template␈α⊂must␈α⊂be␈α∂inferred␈α⊂rather␈α⊂than␈α⊂sought␈α⊂in␈α∂the
␈↓ α←␈↓␈↓ β?text.␈α∀ Such␈α∀inferences␈α∃are␈α∀made␈α∀with␈α∀different␈α∃degrees␈α∀of
␈↓ α←␈↓␈↓ β?confidence.

␈↓ α←␈↓␈↓ ββ(5)␈↓ β?␈↓↓Consistency␈α≤of␈α≤meaning␈α≤assignment␈↓␈α≤(consistency␈α≤of␈α≠data
␈↓ α←␈↓␈↓ β?interpretation).␈α_ Where␈α_ambiguity␈α_exists,␈α_several␈α↔plausible
␈↓ α←␈↓␈↓ β?interpretations␈α≥of␈α≥a␈α≥clause␈α≥may␈α≥arise;␈α≥the␈α≤appropriate
␈↓ α←␈↓␈↓ β?bookkeeping␈αis␈α
done␈αto␈α
assure␈αthat␈α
each␈αword␈α
is␈αunderstood␈α
in
␈↓ α←␈↓␈↓ β?only one way for any given interpretation.

␈↓ α←␈↓␈↓ ββ(6)␈↓ β?␈↓↓Accounting␈α⊃for␈α⊃all␈α⊃words␈α⊃in␈α⊂the␈α⊃text␈↓␈α⊃(accounting␈α⊃for␈α⊃all␈α⊂the
␈↓ α←␈↓␈↓ β?data).␈α∞ Preference␈α∞is␈α∂given␈α∞to␈α∞those␈α∞interpretations␈α∂that␈α∞leave
␈↓ α←␈↓␈↓ β?fewer words unaccounted for in a line of text.

␈↓ α←␈↓␈↓ ββ(7)␈↓ β?␈↓↓Consistency␈α∃of␈α∃attribute,␈α∃object,␈α∃and␈α∃value␈α∃interrelationships␈↓
␈↓ α←␈↓␈↓ β?(internal␈α
consistency␈α
of␈α
hypothesis␈α
structure).␈α
 This␈α
is␈α
used␈α
in
␈↓ α←␈↓␈↓ β?several␈α⊃ways.␈α⊂ For␈α⊃instance,␈α⊂in␈α⊃assembling␈α⊃an␈α⊂interpretation,
␈↓ α←␈↓␈↓ β?ambiguity␈α∪can␈α∀sometimes␈α∪be␈α∀resolved␈α∪simply␈α∀by␈α∪restricting
␈↓ α←␈↓␈↓ β?interpretation␈α∞to␈α∂syntactically␈α∞valid␈α∞triples␈α∂(e.g.,␈α∞there␈α∂may␈α∞be
␈↓ α←␈↓␈↓ β?several␈α↔attributes␈α_and␈α↔values␈α↔suggested␈α_by␈α↔the␈α_text,␈α↔but
␈↓ α←␈↓␈↓ β?perhaps␈α∃only␈α∃one␈α∃of␈α∃the␈α∃possible␈α∃pairings␈α⊗is␈α∃syntactically
␈↓ α←␈↓␈↓ β?valid).␈α⊂ Also,␈α∂knowing␈α⊂one␈α∂of␈α⊂the␈α∂three␈α⊂may␈α∂help␈α⊂guide␈α∂the
␈↓ α←␈↓␈↓ β?search␈α~for␈α→another␈α~(e.g.,␈α→if␈α~the␈α→attribute␈α~is␈α~known␈α→[or
␈↓ α←␈↓␈↓ β?postulated],␈α∩then␈α∪␈↓¬TEIRESIAS␈↓␈α∩can␈α∪refer␈α∩to␈α∪it␈α∩to␈α∪determine␈α∩the
␈↓ α←␈↓␈↓ β?appropriate kinds of values to seek).


␈↓ α←␈↓␈↓ ββ␈↓ β? Expectation-driven knowledge sources:

␈↓ α←␈↓␈↓ ββ␈↓ β?␈↓↓The␈α⊃rule␈α⊃model␈↓.␈α⊃ As␈α⊂noted␈α⊃above,␈α⊃the␈α⊃model␈α⊃chosen␈α⊂during
␈↓ α←␈↓␈↓ β?debugging␈α∞is␈α∞used␈α
as␈α∞a␈α∞source␈α∞of␈α
advice␈α∞(or␈α∞hints)␈α∞about␈α
the
␈↓ α←␈↓␈↓ β?possible content of the new rule.


␈↓ α←␈↓␈↓αTerminology and dictionary structure␈↓    
␈↓ α←␈↓␈↓ β?The␈αEnglish␈αversion␈αof␈αa␈αrule␈αwill␈αbe␈αreferred␈αto␈αas␈α``text,''␈αwhile␈αeach
␈↓ α←␈↓of␈α
the␈α
components␈αof␈α
the␈α
␈↓¬LISP␈↓␈α
version␈αwill␈α
be␈α
referred␈αto␈α
as␈α
a␈α
``␈↓¬LISP␈↓␈αclause.'' 
␈↓ α←␈↓Thus␈α
␈↓↓the␈α
infection␈α∞is␈α
primary-bacteremia␈↓␈α
is␈α∞text,␈α
and␈α
the␈α∞corresponding␈α
␈↓¬LISP␈↓
␈↓ α←␈↓clause␈α_is␈α↔␈↓	(SAME␈α_CNTXT␈α↔INFECTION␈α_PRIMARY-BACTEREMIA)␈↓.␈α_ The␈α↔terms
␈↓ α←␈↓``parse''␈α⊃or␈α⊂``interpretation''␈α⊃are␈α⊂used␈α⊃to␈α⊂mean␈α⊃the␈α⊂creation␈α⊃of␈α⊂a␈α⊃single␈α⊂␈↓¬LISP␈↓
␈↓ α←␈↓premise or action clause.
␈↓ α←␈↓␈↓ β?The␈α
natural-language-understanding␈αcapabilities␈α
are␈α
purely␈αkeyword
␈↓ α←␈↓based.␈α⊃ The␈α⊃connotation␈α⊂of␈α⊃a␈α⊃single␈α⊂word␈α⊃is␈α⊃determined␈α⊂by␈α⊃a␈α⊃number␈α⊂of
␈↓ α←␈↓␈↓5-4␈↓ λ HOW IT ALL WORKS    95␈↓

␈↓"β␈↓ α←␈↓pointers␈α∞associated␈α∞with␈α∞it.␈↓
7␈↓␈α∂These␈α∞are,␈α∞in␈α∞general,␈α∞inverse␈α∂pointers␈α∞derived
␈↓ α←␈↓from␈α∂the␈α∂English␈α∞phrases␈α∂(``translations'')␈α∂associated␈α∞with␈α∂many␈α∂of␈α∂the␈α∞data
␈↓ α←␈↓structures␈α∀in␈α∀the␈α∀system.␈α∀ For␈α∪instance,␈α∀a␈α∀list␈α∀like␈α∀␈↓	STERILESITES␈↓␈α∀has␈α∪a
␈↓ α←␈↓translation␈α⊂of␈α⊂␈↓↓those␈α∂sites␈α⊂which␈α⊂are␈α⊂normally␈α∂sterile.␈↓␈α⊂As␈α⊂a␈α⊂result,␈α∂associated
␈↓ α←␈↓with␈αthe␈αword␈α␈↓↓sterile␈↓␈αis␈α
a␈αpointer␈αto␈αthe␈αname␈α␈↓	STERILESITES␈↓.␈α
 The␈αcreation
␈↓ α←␈↓and␈α∞updating␈α∂of␈α∞these␈α∂pointers␈α∞are␈α∂handled␈α∞semi-automatically␈α∂by␈α∞routines
␈↓ α←␈↓that help to minimize the bookkeeping task.
␈↓ α←␈↓␈↓ β?For␈α
ease␈α
of␈α
reference␈α
later␈α
on,␈α
the␈α
names␈α
of␈α
those␈α
pointers␈α
are␈α
listed
␈↓ α←␈↓below, along with an indication of what they supply.

␈↓"β␈↓ α←␈↓	<WORD1>␈↓

␈↓"β␈↓ α←␈↓␈↓	pointer         value␈↓

␈↓"β␈↓ α←␈↓␈↓	INATTRIB  ␈↓<word1> appears in the translation of these attributes
␈↓"β␈↓ α←␈↓␈↓	INOBJECT  ␈↓<word1> appears in the translation of these object types
␈↓"β␈↓ α←␈↓␈↓	INFUNCS   ␈↓<word1> appears in the translations of these predicate functions
␈↓"β␈↓ α←␈↓␈↓	VALUEOF   ␈↓<word1> is a legal value of these attributes
␈↓"β␈↓ α←␈↓␈↓	INLTRANS  ␈↓<word1> appears in the translation of these list names

␈↓"β␈↓ α←␈↓α␈↓ ∧kFig. 5-4.    Dictionary structure.    

␈↓ α←␈↓They␈α∃are␈α∃referred␈α∃to␈α∃collectively␈α∃as␈α∃``connotation␈α∃pointers.'' ␈α⊗(There␈α∃are
␈↓ α←␈↓additional␈αpointers␈αto␈αhandle␈αsynonyms,␈αbut␈αthey␈αare␈αnot␈αrelevant␈αhere.)␈αThe
␈↓ α←␈↓important␈α∞point␈α∞is␈α∞that␈α
the␈α∞appearance␈α∞of␈α∞any␈α
given␈α∞word␈α∞can␈α∞be␈α∞taken␈α
as
␈↓ α←␈↓evidence␈α⊃that␈α⊃the␈α⊂expert␈α⊃was␈α⊃talking␈α⊃about␈α⊂one␈α⊃of␈α⊃the␈α⊃attributes,␈α⊂objects,
␈↓ α←␈↓predicate␈α∞functions,␈α∞values,␈α∞or␈α∞list␈α∞names␈α∞indicated␈α∞by␈α∞the␈α∞pointers␈α∂for␈α∞that
␈↓ α←␈↓word.

␈↓ α←␈↓␈↓αPre-processing the text␈↓    
␈↓ α←␈↓␈↓ β?The␈α
first␈α
step␈α
in␈α
processing␈α
the␈α
new␈α
rule␈α
is␈α
to␈α
take␈α
a␈α
single␈α
line␈αof␈α
text
␈↓ α←␈↓and␈αreplace␈αeach␈αword␈αwith␈αits␈αroot␈αword␈αequivalent.␈α (Root␈αwords␈αprovide␈αa
␈↓ α←␈↓canonical␈αform␈αfor␈αplurals␈αand␈αother␈αsimple␈αvariations.) ␈αCommon␈αwords␈α
like
␈↓ α←␈↓␈↓↓a,␈α∞and,␈α∞the,␈↓␈α∞are␈α
explicitly␈α∞marked␈α∞in␈α∞the␈α
dictionary␈α∞as␈α∞content␈α∞free,␈α∞and␈α
are
␈↓ α←␈↓ignored.
␈↓ α←␈↓␈↓ β?All␈α
the␈α
connotations␈α
of␈α
each␈α
word␈α
are␈α
then␈α
obtained␈α
by␈α∞referring␈α
to
␈↓ α←␈↓the appropriate pointers.  Fig. 5-5 below shows the result of this process.
␈↓ α←␈↓␈↓ β?As␈α
should␈α
be␈α
clear,␈α
this␈α
technique␈αis␈α
strictly␈α
word␈α
by␈α
word.␈α
 A␈αmore
␈↓ α←␈↓sophisticated␈α∀approach␈α∀would␈α∀have␈α∪some␈α∀grasp␈α∀of␈α∀grammar␈α∀and␈α∪would
␈↓ α←␈↓attempt␈αto␈αassign␈αmeanings␈αto␈αentire␈αphrases.␈α We␈αhave␈αbeen␈αfairly␈α
successful
␈↓ α←␈↓in␈α⊃spite␈α⊃of␈α⊃the␈α⊃primitive␈α⊃approach,␈α⊃primarily␈α⊃because␈α⊃of␈α∩the␈α⊃semi-formal
␈↓ α←␈↓nature␈α∃of␈α∃the␈α⊗technical␈α∃language␈α∃used: ␈α∃It␈α⊗tends␈α∃to␈α∃be␈α⊗terse,␈α∃relatively

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[7]␈α∀The␈α∪current␈α∀structure␈α∪of␈α∀the␈α∪dictionary␈α∀is␈α∪the␈α∀result␈α∪of␈α∀work␈α∀by␈α∪a
␈↓ α←␈↓succession of people associated with the project, in addition to the author.
␈↓ α←␈↓␈↓96    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓unambiguous,␈α
and␈α
has␈α
a␈α
high␈α
information␈α
content.␈α
 ␈↓¬TEIRESIAS␈↓␈α
is␈α
thus␈α
not␈α
often
␈↓ α←␈↓led␈αastray␈αby␈αambiguities␈αor␈α``noise''␈αwords␈αin␈αthe␈αtext.␈α The␈αcommon␈α
lack␈αof
␈↓ α←␈↓correct␈αgrammar␈αin␈αthe␈αexpert's␈αresponses␈αalso␈αsuggests␈αthat␈αany␈αextensive␈αor
␈↓ α←␈↓computationally␈αexpensive␈αuse␈αof␈αa␈αgrammar-based␈αapproach␈αmight␈αnot␈αpay
␈↓ α←␈↓off very well.

␈↓"β␈↓ α←␈↓	TEXT         ROOT WORDS           CONNOTATIONS

␈↓"β␈↓ α←␈↓	the          the                  ␈↓¬NIL␈↓	

␈↓"β␈↓ α←␈↓	patient's    patient              ␈↓¬[NIL (INFECTION) (CURINF) NIL NIL]␈↓	

␈↓"β␈↓ α←␈↓	infection    infection            ␈↓¬[NIL (INFECTION TREATALSO CYTOTOXIC␈↓	
␈↓"β␈↓ α←␈↓	                                  ␈↓¬      NOSOCOMIAL WHENINFECT)␈↓	
␈↓"β␈↓ α←␈↓	                                  ␈↓¬     (CURINF SUSPINF SUSPORG) NIL␈↓	
␈↓"β␈↓ α←␈↓	                                  ␈↓¬     (ALLINFECTIONS)]␈↓	

␈↓"β␈↓ α←␈↓	is           be                   ␈↓¬[(SAME) NIL NIL NIL NIL]␈↓	

␈↓"β␈↓ α←␈↓	primary-     primary-bacteremia   ␈↓¬[NIL NIL NIL (PRIMARY-BACTEREMIA) NIL]␈↓	
␈↓"β␈↓ α←␈↓	bacteremia


␈↓"β
␈↓ α←␈↓	␈↓ β'␈↓αFig.␈α∪5-5.    Original␈α∪text,␈α∪root␈α∪words,␈α∪and␈α∪connotations.  ␈↓The
␈↓ α←␈↓␈↓ β'connotations␈α∞are␈α
listed␈α∞in␈α∞the␈α
order:␈α∞predicate␈α∞function,␈α
attribute,
␈↓ α←␈↓␈↓ β'object,␈α∞value,␈α∂list␈α∞name.␈α∂ Thus,␈α∞␈↓↓patient␈↓␈α∞can␈α∂be␈α∞interpreted␈α∂as␈α∞an
␈↓ α←␈↓␈↓ β'indication␈α
of␈α
either␈α
an␈α
attribute␈α
or␈α
an␈α
object,␈α
while␈α
␈↓↓infection␈↓␈αhas
␈↓ α←␈↓␈↓ β'numerous connotations.


␈↓ α←␈↓␈↓αChecking the rule model␈↓    
␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓'s␈α∂next␈α⊂step␈α∂is␈α∂to␈α⊂verify␈α∂the␈α∂applicability␈α⊂of␈α∂the␈α⊂rule␈α∂model
␈↓ α←␈↓that␈αwas␈α
suggested␈αby␈α
the␈αdebugging␈α
process␈α(or␈α
supply␈αa␈α
model,␈αif␈α
it␈α``came
␈↓ α←␈↓in␈αcold'').␈α This␈αis␈αachieved␈αby␈αscanning␈αeach␈αof␈αthe␈αwords␈αin␈αthe␈αaction␈αpart
␈↓ α←␈↓of␈αthe␈αrule␈αand␈αchecking␈αtheir␈α
connotation␈αpointers␈αto␈αsee␈αwhich␈αattributes␈α
or
␈↓ α←␈↓values are implicated.
␈↓ α←␈↓␈↓ β?If␈α∪the␈α∩debugging␈α∪process␈α∪has␈α∩indicated␈α∪which␈α∩rule␈α∪model␈α∪to␈α∩use,
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓␈αattempts␈α
to␈αreconcile␈αthat␈α
model␈αwith␈αthe␈α
indications␈αfrom␈α
the␈αscan
␈↓ α←␈↓of␈α∩the␈α∪action␈α∩text.␈α∪ If␈α∩there␈α∪is␈α∩agreement,␈α∪nothing␈α∩unusual␈α∪happens.␈α∩ If,
␈↓ α←␈↓however,␈αit␈αcannot␈αreconcile␈αany␈αof␈αthe␈αindications␈αfrom␈αthe␈αaction␈αtext␈αwith
␈↓ α←␈↓the␈α∞given␈α∞model,␈α∞the␈α∞system␈α∞indicates␈α
this␈α∞problem␈α∞and␈α∞asks␈α∞the␈α∞expert␈α
for
␈↓ α←␈↓help.␈α A␈αtrace␈αof␈αthis␈αprocess␈αis␈αshown␈αbelow.␈α Assume␈αthe␈αsystem␈αentered␈αas
␈↓ α←␈↓before,␈α∂expecting␈α∂a␈α∂rule␈α∂about␈α∂category,␈α∂but␈α∂the␈α∂expert␈α∂suddenly␈α⊂thinks␈α∂of
␈↓ α←␈↓another rule he has been meaning to enter:
␈↓ α←␈↓␈↓5-4␈↓ λ HOW IT ALL WORKS    97␈↓


␈↓"β␈↓ α←␈↓	________________________________________
␈↓"β␈↓ α←␈↓	The new rule will be called RULE384
␈↓"β␈↓ α←␈↓	 If     1 -  ␈↓αTHE ORGANISM IS A GRAMNEGATIVE ROD␈↓	
␈↓"β␈↓ α←␈↓	   and  2 -  ␈↓αIT IS ANAEROBIC␈↓	
␈↓"β␈↓ α←␈↓	   and  3 -
␈↓"β␈↓ α←␈↓	 Then:  1 -  ␈↓αTHE ORGANISM MIGHT BE A BACTEROIDES (.6)␈↓	
␈↓"β␈↓ α←␈↓	   and  2 -

␈↓"β␈↓ α←␈↓	Hmm... it seemed that this was going to be a rule to deduce the
␈↓"β␈↓ α←␈↓	category of a organism.   But it doesn't appear to be...

␈↓"β␈↓ α←␈↓	To aid in analyzing this rule, please answer yes or no (Y or N)
␈↓"β␈↓ α←␈↓	to each of the following.

␈↓"β␈↓ α←␈↓	Is this a rule to deduce:
␈↓"β␈↓ α←␈↓	    an organism or class of organisms for which therapy should
␈↓"β␈↓ α←␈↓	    cover
␈↓"β␈↓ α←␈↓	** ␈↓αN␈↓	
␈↓"β␈↓ α←␈↓	    a likely identity of an organism causing a infection
␈↓"β␈↓ α←␈↓	** ␈↓αN␈↓	
␈↓"β␈↓ α←␈↓	    the identity of a organism
␈↓"β␈↓ α←␈↓	** ␈↓αY␈↓	
␈↓"β␈↓ α←␈↓	Thank you...
␈↓"β␈↓ α←␈↓	________________________________________

␈↓ α←␈↓The␈α∞important␈α
thing␈α∞here␈α
is␈α∞a␈α
fast␈α∞and␈α
very␈α∞simple␈α
checking␈α∞process.␈α
 The
␈↓ α←␈↓first␈α∂two␈α∞incorrect␈α∂guesses␈α∞are␈α∂the␈α∞result␈α∂of␈α∞other␈α∂connotations␈α∞of␈α∂the␈α∞word
␈↓ α←␈↓␈↓↓organism␈↓.␈α There␈αare␈αrarely␈αmore␈αthan␈αfour␈αor␈αfive␈αconnotations␈αin␈αany␈αcase,
␈↓ α←␈↓so even at worst, the expert only sees a few bad guesses.
␈↓ α←␈↓␈↓ β?If␈αthe␈αsystem␈αhad␈α``come␈αin␈αcold,''␈αthe␈αprocess␈αwould␈αstart␈αat␈αthe␈αpoint
␈↓ α←␈↓where␈α⊃it␈α⊃says␈α⊃``␈↓	To␈α⊃aid␈α⊃in␈α⊃analyzing....␈↓'' ␈α⊃Even␈α⊃without␈α⊃the␈α⊃debugging
␈↓ α←␈↓process,␈α∞then,␈α∞all␈α∂the␈α∞benefits␈α∞of␈α∂the␈α∞recognition-oriented␈α∞approach␈α∂are␈α∞still
␈↓ α←␈↓available.␈↓
8␈↓
␈↓ α←␈↓␈↓ β?Once␈α∃a␈α∃model␈α∃is␈α∃selected,␈α∃the␈α∃interpretation␈α∃process␈α⊗begins.␈α∃ As
␈↓ α←␈↓indicated,␈αit␈αproceeds␈αline␈αby␈αline␈αand␈αis␈αoriented␈αprimarily␈αtoward␈αdoing␈αthe
␈↓ α←␈↓best␈α∩job␈α∩possible␈α∩with␈α∩the␈α∩limited␈α∩natural␈α∩language␈α∩techniques␈α⊃available.
␈↓ α←␈↓There␈α
are␈α
two␈αimportant␈α
points␈α
to␈αnote.␈α
 First,␈α
the␈αsystem␈α
does␈α
what␈αmight␈α
be
␈↓ α←␈↓called␈α∂``template-directed␈α∞code␈α∂generation,''␈α∞analogous␈α∂to␈α∞the␈α∂way␈α∂English␈α∞is


␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[8]␈αWhile␈α
modeling␈αhuman␈α
performance␈αwas␈αnot␈α
one␈αof␈α
our␈αmotivations,␈αit␈α
is
␈↓ α←␈↓interesting␈α⊗to␈α⊗note␈α⊗that␈α⊗both␈α∃of␈α⊗these␈α⊗reactions␈α⊗are␈α⊗similar␈α⊗to␈α∃human
␈↓ α←␈↓behavior.␈α⊗ Sudden␈α∃changes␈α⊗in␈α⊗the␈α∃topic␈α⊗of␈α∃a␈α⊗conversation␈α⊗can␈α∃violate
␈↓ α←␈↓expectations␈α∞about␈α
what␈α∞is␈α
to␈α∞follow,␈α
resulting␈α∞in␈α
an␈α∞expression␈α∞of␈α
surprise.
␈↓ α←␈↓Similarly,␈αarriving␈αin␈αthe␈αmiddle␈α
of␈αan␈αongoing␈αconversation␈αoften␈αrequires␈α
a
␈↓ α←␈↓moment␈α
to␈α
become␈α
oriented␈α
and␈α
prompts␈α
a␈α
request␈α
for␈α
information.␈α
 Elements
␈↓ α←␈↓of both of these are seen above.
␈↓"β␈↓ α←␈↓␈↓98    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓often␈α
generated␈α
by␈α``filling␈α
in␈α
the␈αblanks''␈α
of␈α
a␈αtemplate.␈↓
9␈↓␈α
Second,␈α
the␈αsystem
␈↓ α←␈↓maintains several types of consistency in the parses that it generates.

␈↓ α←␈↓␈↓αGenerating a LISP clause␈↓    
␈↓ α←␈↓␈↓ β?The␈αnext␈αstep␈αis␈α
to␈αgenerate␈αthe␈αtree␈α
of␈αpossible␈αparses.␈α One␈α
example
␈↓ α←␈↓of␈α⊃the␈α⊃generation␈α⊃process␈α⊃will␈α⊃serve␈α⊃to␈α⊃illustrate␈α⊃the␈α⊃important␈α∩ideas␈α⊃and
␈↓ α←␈↓explain what is meant by a tree of parses.
␈↓ α←␈↓␈↓ β?The␈α∀process␈α∪begins␈α∀by␈α∀determining␈α∪which␈α∀predicate␈α∀function␈α∪the
␈↓ α←␈↓expert␈α∪might␈α∪have␈α∪had␈α∪in␈α∪mind,␈α∪scanning␈α∪the␈α∪list␈α∪of␈α∪connotations,␈α∩and
␈↓ α←␈↓choosing␈α
the␈α
predicate␈α
function␈α
that␈αturns␈α
up␈α
most␈α
often.␈α
 The␈α
template␈αfor
␈↓ α←␈↓this␈α
function␈α
is␈α
retrieved,␈α
and␈α
the␈α
rest␈α
of␈α
the␈α
process␈α
of␈α
creating␈α
a␈α∞clause␈α
is
␈↓ α←␈↓guided␈α⊂by␈α⊂the␈α∂attempt␈α⊂to␈α⊂fill␈α∂in␈α⊂the␈α⊂template.␈↓
10␈↓␈α∂For␈α⊂example,␈α⊂suppose␈α∂the
␈↓ α←␈↓function is ␈↓	SAME␈↓ and the template is

␈↓"β␈↓ α←␈↓	␈↓ ¬∂(SAME CNTXT ATTRIB VALUE)

␈↓ α←␈↓␈↓ β?Associated␈αwith␈αeach␈αof␈αthe␈αprimitives␈αin␈αthe␈αhigh-level␈α``language''␈α
is
␈↓ α←␈↓a␈α∪routine␈α∪that␈α∪embodies␈α∪much␈α∪of␈α∪the␈α∪semantics␈α∪of␈α∪each␈α∀primitive.␈α∪The
␈↓ α←␈↓template␈α⊂is␈α⊂filled␈α⊂in␈α⊂by␈α⊂calling␈α⊂each␈α⊂routine␈α⊂as␈α⊂needed␈α⊂and␈α⊂allowing␈α⊂it␈α⊂to
␈↓ α←␈↓examine␈α∪the␈α∪list␈α∪of␈α∪connotations␈α∪to␈α∪find␈α∪the␈α∪kind␈α∪of␈α∪object␈α∪it␈α∪requires.
␈↓ α←␈↓Consider␈α∂the␈α∂text␈α∂from␈α∞the␈α∂first␈α∂line␈α∂in␈α∞the␈α∂trace,␈α∂␈↓↓the␈α∂patient's␈α∂infection␈α∞is
␈↓ α←␈↓↓primary-bacteremia␈↓.␈α⊃ The␈α⊃routine␈α⊃for␈α⊃␈↓	VALUE␈↓␈α⊃would␈α⊃discover␈α⊃that␈α∩the␈α⊃only
␈↓ α←␈↓object␈α
of␈α
type␈α␈↓	VALUE␈↓␈α
suggested␈α
by␈αthe␈α
text␈α
was␈α
␈↓	PRIMARY-BACTEREMIA␈↓.␈α The
␈↓ α←␈↓routine␈α
for␈α
␈↓	ATTRIB␈↓␈α
would␈α
find␈α
several␈α
objects␈α
of␈α
type␈α
␈↓	ATTRIBUTE␈↓,␈α
since␈αthe
␈↓ α←␈↓word␈α∪␈↓↓infection␈↓␈α∩has␈α∪a␈α∩large␈α∪number␈α∩of␈α∪connotations.␈α∩ However,␈α∪since␈α∩the
␈↓ α←␈↓␈↓	VALUE␈↓␈α∞routine␈α∂has␈α∞already␈α∂filled␈α∞in␈α∞␈↓	PRIMARY-BACTEREMIA␈↓␈α∂as␈α∞a␈α∂␈↓	VALUE␈↓,␈α∞this
␈↓ α←␈↓narrows␈α∞the␈α
choice␈α∞to␈α∞those␈α
attributes␈α∞for␈α
which␈α∞␈↓	PRIMARY-BACTEREMIA␈↓␈α∞is␈α
a
␈↓ α←␈↓legal␈α⊂value,␈α⊃and␈α⊂the␈α⊂routine␈α⊃takes␈α⊂the␈α⊂first␈α⊃of␈α⊂them␈α⊃(␈↓	INFECTION␈↓)␈α⊂initially.
␈↓ α←␈↓The␈αroutine␈αfor␈α␈↓	CNTXT␈↓␈αnotices␈αthat␈α␈↓↓patient␈↓␈αcan␈αbe␈αinterpreted␈αas␈αan␈αobject,␈↓
11␈↓
␈↓ α←␈↓one␈αthat␈αis␈αa␈αvalid␈αobject␈αfor␈α
the␈αattribute␈α␈↓	INFECTION␈↓,␈αand␈αhence␈αmarks␈αit␈α
as
␈↓ α←␈↓such.␈α
 It␈α
returns␈α
the␈α
literal␈α∞atom␈α
␈↓	CNTXT␈↓,␈α
reflecting␈α
the␈α
fact␈α
that␈α∞␈↓	CNTXT␈↓␈α
plays
␈↓ α←␈↓the role of a free variable that is bound when the rule is invoked.
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[9]␈αThe␈αbasic␈αconcept␈αof␈αfilling␈αin␈αa␈αtemplate␈αto␈αgenerate␈αcode␈αis␈αtaken␈αfrom
␈↓ α←␈↓the␈α
original␈α
design␈α
in␈α
[Shortliffe76].␈α
 All␈α
the␈α
rest␈α
of␈α
the␈α
process␈α
that␈α
guides
␈↓ α←␈↓the␈α∞template␈α∞completion␈α∞was␈α∞designed␈α
by␈α∞the␈α∞author,␈α∞including␈α∞the␈α∞view␈α
of
␈↓ α←␈↓attributes, values, etc., as extended data types.

␈↓ α←␈↓[10]␈α
Note␈α
that␈α
this␈α
is␈α
the␈α
same␈α
template␈α
described␈α
in␈α
chapter␈α
2␈α
as␈α
the␈αbasis
␈↓ α←␈↓for␈α
the␈α
system's␈α
ability␈αto␈α
dissect␈α
a␈α
rule.␈α
 It␈αhas␈α
recently␈α
been␈α
pointed␈α
out␈αto
␈↓ α←␈↓me␈αthat␈αthe␈αtechniques␈αused␈αto␈αfill␈αin␈αthe␈αtemplate␈αare␈αsimilar␈αto␈αsome␈αof␈αthe
␈↓ α←␈↓work␈α≥on␈α≥natural␈α≥language␈α≥parsing␈α≥using␈α≥case␈α≥grammars␈α≡(see,␈α≥e.g.,
␈↓ α←␈↓[Rumelhart73]).

␈↓ α←␈↓[11]␈α∪Recall␈α∪that␈α∀objects␈α∪are␈α∪also␈α∪referred␈α∀to␈α∪as␈α∪``contexts,''␈α∀for␈α∪historical
␈↓ α←␈↓reasons.
␈↓ α←␈↓␈↓5-4␈↓ λ HOW IT ALL WORKS    99␈↓

␈↓"β␈↓ α←␈↓␈↓ β?This produces the clause ␈↓
12␈↓

␈↓"β␈↓ α←␈↓	␈↓ ∧∂(SAME CNTXT INFECTION PRIMARY-BACTEREMIA)

␈↓ α←␈↓All␈α∂the␈α∞nontrivial␈α∂words␈α∂in␈α∞the␈α∂text␈α∞have␈α∂been␈α∂assigned␈α∞a␈α∂meaning,␈α∂so␈α∞no
␈↓ α←␈↓more␈α→clauses␈α→can␈α→be␈α→derived␈α_from␈α→it.␈α→There␈α→are,␈α→however,␈α_alternate
␈↓ α←␈↓interpretations␈αfor␈αtwo␈αof␈αthe␈αwords␈α(␈↓↓patient␈↓␈αand␈α␈↓↓infection␈↓).␈α The␈αsystem␈αuses
␈↓ α←␈↓a␈αstandard␈αdepth-first␈αsearch␈αwith␈α
backtracking␈αand,␈αat␈αthis␈αpoint,␈αundoes␈α
its
␈↓ α←␈↓current␈α∩set␈α∩of␈α∩meaning␈α∩assignments␈α∩and␈α∩tries␈α∩all␈α∩the␈α∩alternatives.␈α⊃ Other
␈↓ α←␈↓clauses are generated as alternate interpretations.
␈↓ α←␈↓␈↓ β?It␈αwill␈αbe␈αinstructive␈αto␈αexamine␈αone␈αtype␈αof␈αfailure␈αthat␈αcan␈αoccur␈αas
␈↓ α←␈↓clauses␈α
are␈α
generated.␈α
 This␈αwill␈α
illustrate␈α
the␈α
point␈αmade␈α
in␈α
chapter␈α
2,␈αthat
␈↓ α←␈↓knowledge␈αacquisition␈αis␈αbased␈αin␈αpart␈αon␈αgiving␈αthe␈αsystem␈αaccess␈αto␈αand␈αan
␈↓ α←␈↓``understanding''␈αof␈αits␈αown␈αrepresentations.␈α To␈αsee␈αthis,␈αconsider␈αthe␈αsystem's
␈↓ α←␈↓response␈α∩to␈α∩a␈α∪line␈α∩like␈α∩␈↓↓the␈α∪site␈α∩of␈α∩the␈α∩culture␈α∪is␈α∩blood␈↓.␈α∩ The␈α∪system␈α∩will
␈↓ α←␈↓generate␈α∞several␈α∂clauses,␈α∞including␈α∂the␈α∞correct␈α∂one␈α∞(which␈α∂uses␈α∞␈↓	SAME␈↓␈α∂as␈α∞the
␈↓ α←␈↓predicate␈α∞function␈α∞and␈α
interprets␈α∞␈↓↓culture␈↓␈α∞as␈α∞the␈α
object,␈α∞␈↓↓site␈↓␈α∞as␈α∞the␈α
attribute,
␈↓ α←␈↓and␈α␈↓↓blood␈↓␈α
as␈αthe␈α
value).␈α In␈α
one␈αof␈α
its␈αlater␈α
attempts␈αto␈α
construct␈αother␈α
clauses,
␈↓ α←␈↓it␈αwill␈αdiscover␈αit␈αhas␈αused␈αall␈αthe␈α␈↓	VALUE␈↓s␈αit␈αcan␈αfind,␈αand␈αthe␈α␈↓	VALUE␈↓␈αroutine
␈↓ α←␈↓thus␈αfails.␈α
 The␈αroutine␈α
for␈α␈↓	ATTRIB␈↓␈α
then␈αfinds␈α
it␈αhas␈α
not␈αyet␈α
tried␈αto␈α
interpret
␈↓ α←␈↓␈↓↓culture␈↓␈α∪as␈α∪an␈α∪indication␈α∪of␈α∀the␈α∪attribute␈α∪``␈↓	the␈α∪number␈α∪of␈α∀cultures␈α∪in
␈↓ α←␈↓	this␈α⊂series␈↓''␈α⊂(␈↓	NUMCULS␈↓)␈α⊂and␈α⊂makes␈α⊂this␈α⊂assignment.␈α⊂ It␈α⊂then␈α⊂invokes␈α∂the
␈↓ α←␈↓␈↓	VALUE␈↓␈α⊃routine␈α⊂with␈α⊃instructions␈α⊃to␈α⊂look␈α⊃for␈α⊃a␈α⊂value␈α⊃for␈α⊃␈↓	NUMCULS␈↓.␈↓
13␈↓␈α⊂That
␈↓ α←␈↓routine,␈α∀in␈α∀turn,␈α∀uses␈α∪its␈α∀knowledge␈α∀of␈α∀the␈α∪structure␈α∀of␈α∀an␈α∀attribute␈α∪to
␈↓ α←␈↓examine␈α
␈↓	NUMCULS␈↓,␈α∞discovers␈α
that␈α
it␈α∞takes␈α
an␈α∞integer␈α
between␈α
1␈α∞and␈α
15␈α∞as␈α
a
␈↓ α←␈↓value,␈α∂and␈α∂finds␈α∂none␈α∂in␈α∂the␈α∂text.␈α∂ The␈α∂attempt␈α∂to␈α∂interpret␈α∂␈↓↓culture␈↓␈α∂as␈α∂an
␈↓ α←␈↓indication␈α⊂of␈α⊂␈↓	NUMCULS␈↓␈α⊂thus␈α⊂fails,␈α⊃the␈α⊂assignment␈α⊂is␈α⊂undone,␈α⊂and␈α⊃the␈α⊂next
␈↓ α←␈↓alternative␈α∀is␈α∀tried.␈α∀ Maintaining␈α∀the␈α∀internal␈α∀consistency␈α∀of␈α∃the␈α∀clauses
␈↓ α←␈↓generated␈α
is␈α
thus␈α
based␈α
in␈α
part␈α
on␈α
giving␈α
the␈α
system␈α
the␈α
ability␈α
to␈αexamine
␈↓ α←␈↓its␈α⊂own␈α⊂representations;␈α⊂in␈α⊂this␈α⊂case,␈α⊂an␈α⊂attribute␈α⊂is␈α⊂examined␈α⊂to␈α⊂find␈α∂out
␈↓ α←␈↓what kind of values are associated with it.
␈↓ α←␈↓␈↓ β?As␈α∞may␈α∞be␈α∞clear,␈α∞the␈α∞template␈α∞parts␈α∞are␈α∞not␈α∞necessarily␈α∞filled␈α∞in␈α∞the
␈↓ α←␈↓order␈αin␈α
which␈αthey␈α
appear␈αin␈α
the␈αtemplate.␈α
 The␈α␈↓	VALUE␈↓␈α
part␈α(if␈α
there␈αis␈α
one)
␈↓ α←␈↓is␈α
always␈α
tried␈α
first,␈α
for␈α
instance,␈α
since␈α
words␈α
indicating␈α
values␈α
in␈αthis␈α
domain

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[12]␈α⊃Since␈α∩the␈α⊃clauses␈α⊃are␈α∩later␈α⊃scored␈α⊃for␈α∩likely␈α⊃validity,␈α⊃the␈α∩first␈α⊃clause
␈↓ α←␈↓generated is not necessarily the system's first guess.

␈↓ α←␈↓[13]␈α∪The␈α∪␈↓	VALUE␈↓␈α∪routine␈α∪can␈α∪be␈α∪re-invoked,␈α∪despite␈α∪its␈α∪previous␈α∩failure,
␈↓ α←␈↓because␈α
there␈α∞are␈α
a␈α∞number␈α
of␈α∞cases␈α
in␈α∞which␈α
the␈α∞word-by-word␈α
approach
␈↓ α←␈↓fails.␈α It␈αis␈αthen␈αprofitable␈αto␈αgo␈αback␈αand␈αtake␈αanother␈αlook␈αas␈αlong␈αas␈αthere
␈↓ α←␈↓is␈α
some␈α
idea␈α
of␈α
what␈α
to␈α
look␈α
for.␈α
 For␈α
attributes␈α
which␈α
are␈α
either␈α
true␈α
or␈α
false,
␈↓ α←␈↓for␈α
instance,␈αan␈α
indication␈αof␈α
the␈α␈↓	VALUE␈↓␈α
rarely␈αappears␈α
explicitly␈αin␈α
the␈αtext
␈↓ α←␈↓(e.g.,␈αthe␈α
text␈αmight␈α
be␈α␈↓↓the␈α
patient␈αis␈α
a␈αcompromised␈α
host␈↓␈αrather␈α
than␈α␈↓↓it␈αis␈α
true
␈↓ α←␈↓↓that the patient is a compromised host␈↓).
␈↓ α←␈↓␈↓100    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓are␈αoften␈αtotally␈αunambiguous.␈↓
14␈↓␈αThis␈αsimple␈α``first␈αpass''␈αstrategy␈αis␈αbuilt␈α
into
␈↓ α←␈↓the␈α
driver␈αroutines,␈α
but,␈α
since␈αone␈α
routine␈α
may␈αinvoke␈α
another,␈α
the␈αorder␈α
may
␈↓ α←␈↓soon become more complex.
␈↓ α←␈↓␈↓ β?The␈α⊂entire␈α⊂tree␈α⊂of␈α⊂parses␈α⊂is␈α⊂generated␈α⊂using␈α⊂the␈α⊂depth-first␈α∂search
␈↓ α←␈↓with␈α∂the␈α⊂backup␈α∂noted␈α⊂earlier.␈α∂ The␈α⊂result␈α∂is␈α∂a␈α⊂tree␈α∂of␈α⊂clauses␈α∂of␈α⊂the␈α∂sort
␈↓ α←␈↓shown␈αbelow␈α(part␈αof␈αthe␈αtree␈αgenerated␈αfrom␈αthe␈αfirst␈αline␈αof␈αtext␈αis␈αshown).
␈↓ α←␈↓At␈α
each␈α∞node␈α
of␈α
the␈α∞tree␈α
is␈α
a␈α∞potential␈α
premise␈α
clause,␈α∞and␈α
any␈α∞single␈α
path
␈↓ α←␈↓through the tree from the root to a leaf is a set of consistent interpretations.

␈↓"β␈↓ α←␈↓¬                                          /\
␈↓"β␈↓ α←␈↓¬                                        /    \
␈↓"β␈↓ α←␈↓¬                                      /        \
␈↓"β␈↓ α←␈↓¬                                    /            \
␈↓"β␈↓ α←␈↓¬                                  /                \
␈↓"β␈↓ α←␈↓¬                                /                    \

␈↓"β␈↓ α←␈↓¬(SAME CNTXT INFECTION PRIMARY-BACTEREMIA)  (SAME CNTXT TREATALSO PRIMARY-BACTEREMIA)

␈↓"β␈↓ α←␈↓¬                                                 /       \
␈↓"β␈↓ α←␈↓¬                                                /         \
␈↓"β␈↓ α←␈↓¬                                               /           \
␈↓"β␈↓ α←␈↓¬                                              /             \
␈↓"β␈↓ α←␈↓¬                                             /               \

␈↓"β␈↓ α←␈↓¬                           (SAME CNTXT CYTOTOXIC)           (SAME CNTXT NOSOCOMIAL)

␈↓ α←␈↓␈↓ β?Generation␈α∂of␈α∂the␈α∂tree␈α⊂with␈α∂alternative␈α∂word␈α∂meanings␈α⊂in␈α∂different
␈↓ α←␈↓branches␈α≤provides␈α≤the␈α≥notion␈α≤of␈α≤consistency␈α≥between␈α≤interpretations.
␈↓ α←␈↓Consistency␈αis␈αrequired␈αin␈α
order␈αto␈αpermit␈α␈↓¬TEIRESIAS␈↓␈α
to␈αget␈αmore␈αthan␈αone␈α
␈↓¬LISP␈↓
␈↓ α←␈↓premise␈α↔clause␈α_from␈α↔a␈α_single␈α↔line␈α↔of␈α_text␈α↔without␈α_making␈α↔conflicting
␈↓ α←␈↓assumptions␈α∞about␈α∞that␈α
text.␈α∞ The␈α∞current␈α
implementation␈α∞takes␈α∞care␈α∞of␈α
the
␈↓ α←␈↓most␈α
obvious␈αsources␈α
of␈αsuch␈α
conflicts␈α
by␈αinsuring␈α
that␈αonly␈α
one␈α
meaning␈αis
␈↓ α←␈↓chosen for each word.
␈↓ α←␈↓␈↓ β?There␈α∞are␈α∞a␈α
number␈α∞of␈α∞factors␈α
that␈α∞prevent␈α∞the␈α
tree␈α∞of␈α∞parses␈α
from
␈↓ α←␈↓becoming␈α∞unreasonably␈α
large.␈α∞ To␈α∞explain␈α
the␈α∞first␈α∞factor,␈α
it␈α∞is␈α∞necessary␈α
to
␈↓ α←␈↓refine␈α∪the␈α∪statement␈α∀above␈α∪which␈α∪claimed␈α∀that,␈α∪in␈α∪choosing␈α∀a␈α∪predicate
␈↓ α←␈↓function,␈α∂the␈α∂first␈α∂one␈α∂chosen␈α∂is␈α∂``the␈α∂one␈α∂that␈α∂turns␈α∂up␈α∂most␈α∂often.''␈α∂More
␈↓ α←␈↓precisely,␈α∞when␈α∞choosing␈α∞a␈α
predicate␈α∞function␈α∞(or␈α∞any␈α
item␈α∞to␈α∞fill␈α∞in␈α∞one␈α
of
␈↓ α←␈↓the␈α
blanks),␈αthe␈α
first␈α
one␈αchosen␈α
is␈α
``the␈αone␈α
that␈α
turns␈αup␈α
most␈α
often,␈α␈↓↓and␈↓␈α
was
␈↓ α←␈↓suggested␈αby␈α
unambiguous␈αwords.'' ␈αThis␈α
means␈αthat␈αthe␈α
tree␈αis␈αgenerated␈α
far
␈↓ α←␈↓more␈α∩efficiently,␈α∩with␈α∪less␈α∩backup.␈α∩ An␈α∩example␈α∪is␈α∩shown␈α∩below,␈α∪in␈α∩two
␈↓ α←␈↓versions␈α∂of␈α∞a␈α∂fragment␈α∞of␈α∂a␈α∞parse␈α∂tree␈α∞that␈α∂would␈α∞result␈α∂from␈α∞the␈α∂text␈α∞␈↓↓the

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[14]␈αThey␈αare␈αalso␈αsufficiently␈αstrong␈αclues␈αas␈αto␈αtext␈αcontent␈αthat␈αthe␈α␈↓	ATTRIB␈↓
␈↓ α←␈↓routines␈αwill␈αsupply␈αthe␈αattribute␈αwhich␈αbelongs␈αwith␈αa␈αgiven␈α␈↓	VALUE␈↓,␈αeven␈αif
␈↓ α←␈↓no␈α
other␈αindication␈α
of␈α
that␈αattribute␈α
can␈αbe␈α
found.␈α
 This␈αis␈α
what␈α
allows␈αthe
␈↓ α←␈↓system␈α⊃to␈α∩parse␈α⊃something␈α∩like␈α⊃␈↓↓the␈α∩organism␈α⊃is␈α∩a␈α⊃gramnegative␈α∩rod␈↓,␈α⊃even
␈↓ α←␈↓though␈α⊂there␈α⊂are␈α∂no␈α⊂other␈α⊂indicators␈α∂of␈α⊂gram␈α⊂stain␈α∂(other␈α⊂than␈α⊂the␈α∂value
␈↓ α←␈↓␈↓↓gramnegative␈↓) or morphology (other than the value ␈↓↓rod␈↓).
␈↓"β␈↓ α←␈↓␈↓5-4␈↓ λ⊃HOW IT ALL WORKS    101␈↓

␈↓"β␈↓ α←␈↓↓organism␈α∃is␈α∀an␈α∃aerobic␈α∀rod␈↓.␈α∃ ␈↓↓Aerobic␈↓␈α∀is␈α∃ambiguous␈α∀(it␈α∃can␈α∀be␈α∃either␈α∀an
␈↓ α←␈↓aerobicity␈α∩value␈α∩or␈α∪an␈α∩organism␈α∩subtype),␈α∪while␈α∩␈↓↓rod␈↓␈α∩is␈α∪unambiguously␈α∩a
␈↓ α←␈↓morphology value.

␈↓"␈↓ α←␈↓∧                                /   \
␈↓"␈↓ α←␈↓∧                               /     \
␈↓"␈↓ α←␈↓∧                              /       \

␈↓"␈↓ α←␈↓∧      (SAME CNTXT SUBTYPE AEROBIC)  (SAME CNTXT AIR AEROBIC)

␈↓"␈↓ α←␈↓∧                     |                        |
␈↓"␈↓ α←␈↓∧                     |                        |

␈↓"␈↓ α←␈↓∧          (SAME CNTXT MORPH ROD)      (SAME CNTXT MORPH ROD)


␈↓"␈↓ α←␈↓α␈↓ ∧∩Fig. 5-6.    Inefficient tree of interpretations.    


␈↓"␈↓ α←␈↓∧                                 |
␈↓"␈↓ α←␈↓∧                                 |
␈↓"␈↓ α←␈↓∧                                 |

␈↓"␈↓ α←␈↓∧                      (SAME CNTXT MORPH ROD)

␈↓"␈↓ α←␈↓∧                             /      \
␈↓"␈↓ α←␈↓∧                            /        \
␈↓"␈↓ α←␈↓∧                           /          \

␈↓"␈↓ α←␈↓∧      (SAME CNTXT SUBTYPE AEROBIC)  (SAME CNTXT AIR AEROBIC)


␈↓"␈↓ α←␈↓α␈↓ β{Fig. 5-7.    More efficient tree of interpretations.    

␈↓ α←␈↓␈↓ β?Second,␈α
the␈αinsistence␈α
on␈αconsistency␈α
within␈αa␈α
parse␈αis␈α
simple␈αbut␈α
very
␈↓ α←␈↓effective.␈α∩ It␈α⊃is␈α∩more␈α∩efficient␈α⊃to␈α∩avoid␈α∩producing␈α⊃invalid␈α∩parses␈α∩than␈α⊃to
␈↓ α←␈↓generate them and prune them out later.
␈↓ α←␈↓␈↓ β?Finally,␈αand␈αperhaps␈αmost␈αimportant,␈αthere␈αis␈αa␈αvery␈αsmall␈αamount␈αof
␈↓ α←␈↓text␈αand␈αmost␈αof␈αit␈αis␈αrelatively␈αunambiguous.␈α While␈αthe␈αexpert␈αis␈αpermitted
␈↓ α←␈↓to␈α∂type␈α∂an␈α∂arbitrary␈α∂amount␈α∂in␈α∂each␈α∂text␈α∂line,␈α∂there␈α∂are␈α∂typically␈α∂no␈α∂more
␈↓ α←␈↓than 15 or so words.
␈↓ α←␈↓␈↓ β?At␈αthis␈αpoint␈α
the␈αsystem␈αhas␈αgenerated␈α
a␈αtree␈αof␈αinterpretations␈α
for␈αa
␈↓ α←␈↓single␈α⊃line␈α⊂of␈α⊃text.␈α⊃ Any␈α⊂path␈α⊃in␈α⊃that␈α⊂tree␈α⊃from␈α⊃root␈α⊂to␈α⊃leaf␈α⊃represents␈α⊂a
␈↓ α←␈↓consistent␈α∂and␈α∞plausible␈α∂set␈α∞of␈α∂␈↓¬LISP␈↓␈α∂clauses␈α∞that␈α∂contain␈α∞all␈α∂of␈α∂the␈α∞meaning
␈↓ α←␈↓that could be found in the text.
␈↓ α←␈↓␈↓ β?As␈α∩each␈α∩clause␈α∪is␈α∩completed,␈α∩it␈α∩is␈α∪given␈α∩a␈α∩preliminary␈α∪score␈α∩that
␈↓ α←␈↓reflects␈α↔how␈α↔it␈α↔was␈α_assembled.␈α↔ For␈α↔instance,␈α↔those␈α↔clauses␈α_for␈α↔which
␈↓ α←␈↓independent␈α
evidence␈α
was␈α
found␈α
for␈αboth␈α
␈↓	VALUE␈↓␈α
and␈α
␈↓	ATTRIB␈↓␈α
are␈α
given␈αthe
␈↓ α←␈↓highest␈αscore.␈α If␈αthe␈α␈↓	ATTRIB␈↓␈αmust␈αbe␈αimplied␈αby␈αthe␈αpresence␈αof␈αa␈αvalue,␈αthe
␈↓ α←␈↓␈↓102    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓score␈αis␈αsomewhat␈αlower.␈α There␈αare␈αa␈αnumber␈αof␈αother␈αcriteria␈αthat␈αare␈αalso
␈↓ α←␈↓used␈α⊃in␈α⊃composing␈α∩the␈α⊃score.␈α⊃ The␈α⊃main␈α∩idea␈α⊃is␈α⊃simply␈α⊃to␈α∩provide␈α⊃some
␈↓ α←␈↓measure␈αof␈αhow␈αstrongly␈αthe␈αdata␈α(the␈αtext)␈αsuggested␈αthe␈αinterpretations␈α(the
␈↓ α←␈↓␈↓¬LISP␈↓ clauses) that were made.

␈↓ α←␈↓␈↓αScoring the parses␈↓    
␈↓ α←␈↓␈↓ β?The␈α∩next␈α∪step␈α∩is␈α∪to␈α∩select␈α∩a␈α∪single␈α∩interpretation␈α∪for␈α∩the␈α∪text␈α∩by
␈↓ α←␈↓choosing␈αa␈α
path␈αthrough␈α
the␈αtree␈α
of␈αclauses.␈α
 This␈αis␈α
done␈αwith␈α
reference␈αto
␈↓ α←␈↓the␈α⊂rule␈α⊂model␈α⊃chosen␈α⊂during␈α⊂the␈α⊂debugging␈α⊃phase.␈α⊂ Each␈α⊂path␈α⊃is␈α⊂scored
␈↓ α←␈↓according␈α⊂to␈α⊃how␈α⊂well␈α⊂it␈α⊃fulfills␈α⊂the␈α⊂expectations␈α⊃expressed␈α⊂by␈α⊃the␈α⊂model.
␈↓ α←␈↓The␈α∀singlets␈α∪in␈α∀the␈α∪model␈α∀predict␈α∪the␈α∀appearance␈α∪of␈α∀clauses␈α∪containing
␈↓ α←␈↓specific␈α∪attributes␈α∪and␈α∪predicate␈α∀functions,␈α∪while␈α∪the␈α∪ntuples␈α∀predict␈α∪the
␈↓ α←␈↓appearance␈αof␈αassociations␈αof␈αclauses␈αcontaining␈αcertain␈αattributes.␈α The␈α
score
␈↓ α←␈↓for each path is the sum of the strengths of the predictions that it fulfills.
␈↓ α←␈↓␈↓ β?There␈α⊂are␈α⊂thus␈α⊃two␈α⊂scores.␈α⊂ The␈α⊂individual␈α⊃score␈α⊂for␈α⊂a␈α⊃given␈α⊂␈↓¬LISP␈↓
␈↓ α←␈↓clause␈α∞indicates␈α∂how␈α∞strongly␈α∞the␈α∂clause␈α∞was␈α∞suggested␈α∂by␈α∞the␈α∂English␈α∞text.
␈↓ α←␈↓The␈α⊃score␈α⊃for␈α⊂an␈α⊃entire␈α⊃path␈α⊂indicates␈α⊃how␈α⊃well␈α⊂the␈α⊃set␈α⊃of␈α⊃clauses␈α⊂meets
␈↓ α←␈↓expectations.␈α⊗ These␈α⊗two␈α⊗are␈α↔combined␈α⊗in␈α⊗a␈α⊗way␈α⊗that␈α↔emphasizes␈α⊗the
␈↓ α←␈↓expectations␈α
(the␈α
recognition-oriented␈α
approach),␈α
and␈α
candidates␈α
are␈αranked
␈↓ α←␈↓according␈α⊂to␈α⊃the␈α⊂outcome.␈α⊃ The␈α⊂system␈α⊂will␈α⊃thus␈α⊂``hear␈α⊃what␈α⊂it␈α⊃expects␈α⊂to
␈↓ α←␈↓hear''␈α∞if␈α∞that␈α∞is␈α∞at␈α∞all␈α∞possible;␈α∞otherwise,␈α∞it␈α∞will␈α∞choose␈α∞the␈α∂best␈α∞alternative
␈↓ α←␈↓interpretation.

␈↓ α←␈↓␈↓α5-4-3    Checking results␈↓
␈↓ α←␈↓␈↓ ¬GMay I say what I think second best?
␈↓ α←␈↓␈↓ ¬GIf there's a third best, too, spare not to tell it.
␈↓"β␈↓ α←␈↓␈↓ π+␈↓↓Oedipus the King␈↓, lines 282-283
␈↓ α←␈↓␈↓ β?Having␈α∞chosen␈α∞a␈α
likely␈α∞interpretation␈α∞for␈α∞each␈α
line␈α∞of␈α∞text␈α∞typed␈α
by
␈↓ α←␈↓the expert, the system displays the entire rule and asks for approval.

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬This is my understanding of your rule:
␈↓"β␈↓ α←␈↓¬RULE383
␈↓"β␈↓ α←␈↓¬-------
␈↓"β␈↓ α←␈↓¬  If:  1) Primary-bacteremia is an associated infection for which therapy should
␈↓"β␈↓ α←␈↓¬          cover, and
␈↓"β␈↓ α←␈↓¬       2) The infection with the organism was acquired while the patient was
␈↓"β␈↓ α←␈↓¬          hospitalized, and
␈↓"β␈↓ α←␈↓¬       3) The culture was taken from a sterile source, and
␈↓"β␈↓ α←␈↓¬       4) The organism is not normally found at this site
␈↓"β␈↓ α←␈↓¬  Then:  There is strongly suggestive evidence (.8) that the category of the
␈↓"β␈↓ α←␈↓¬         organism is enterobacteriaceae

␈↓"β␈↓ α←␈↓¬Okay? (Yes or No)
␈↓"β␈↓ α←␈↓¬--**␈↓π N
␈↓"β␈↓ α←␈↓π_______________________________________

␈↓ α←␈↓␈↓ β?When␈α
the␈α
expert␈αindicates␈α
a␈α
problem,␈α
the␈αsystem␈α
invokes␈α
a␈αrule␈α
editor
␈↓ α←␈↓that␈αallows␈α
him␈αto␈αmake␈α
changes.␈α He␈α
can␈αadd␈αa␈α
clause␈α(the␈α
system␈αprompts
␈↓ α←␈↓for␈α
a␈α∞new␈α
line␈α∞of␈α
text),␈α∞delete␈α
one,␈α
undo␈α∞a␈α
deletion␈α∞(``undeleting''),␈α
or␈α∞ask␈α
to
␈↓ α←␈↓change a clause.
␈↓"β␈↓ α←␈↓␈↓5-4␈↓ λ⊃HOW IT ALL WORKS    103␈↓

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬Change premise?
␈↓"β␈↓ α←␈↓¬--**␈↓π Y␈↓¬

␈↓"β␈↓ α←␈↓¬What should be changed in premise? (type HELP for list of commands)
␈↓"β␈↓ α←␈↓¬++**␈↓π CHANGE 1␈↓¬

␈↓"β␈↓ α←␈↓¬Suppose (1) were changed to read-
␈↓"β␈↓ α←␈↓¬        1 - The infection is primary-bacteremia
␈↓"β␈↓ α←␈↓¬Would that be ok?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬

␈↓"β␈↓ α←␈↓¬Then you'll probably want to delete:
␈↓"β␈↓ α←␈↓¬       2) The infection with the organism was acquired while the patient was
␈↓"β␈↓ α←␈↓¬          hospitalized.
␈↓"β␈↓ α←␈↓¬Correct?
␈↓"β␈↓ α←␈↓¬++** ␈↓πY
␈↓"β␈↓ α←␈↓π_______________________________________

␈↓ α←␈↓␈↓ β?The␈α``change''␈αcommand␈αmakes␈αuse␈αof␈αthe␈αfact␈αthat␈αall␈αthe␈αalternative
␈↓ α←␈↓interpretations␈α⊃of␈α⊃each␈α⊃text␈α⊃line␈α⊃have␈α⊃been␈α⊃saved.␈α⊃ Given␈α⊃a␈α⊃clause␈α∩to␈α⊃be
␈↓ α←␈↓changed,␈α∂the␈α∂system␈α∂invokes␈α∂several␈α∂heuristics␈α∂to␈α∂find␈α∂the␈α∂next␈α∂most␈α∞likely
␈↓ α←␈↓interpretation.␈α↔ The␈α↔primary␈α↔heuristic␈α↔is␈α⊗to␈α↔examine␈α↔the␈α↔clause␈α↔for␈α⊗a
␈↓ α←␈↓component␈α(a␈α
predicate␈αfunction,␈α
attribute,␈αvalue,␈αetc.)␈α
that␈αwas␈α
suggested␈αby
␈↓ α←␈↓an␈α∞ambiguous␈α
word␈α∞in␈α
the␈α∞text,␈α
to␈α∞see␈α
if␈α∞there␈α
is␈α∞an␈α
alternative␈α∞clause␈α
that
␈↓ α←␈↓uses␈α∂one␈α∞of␈α∂the␈α∂other␈α∞connotations␈α∂of␈α∞that␈α∂word.␈α∂ The␈α∞system␈α∂thus␈α∂tries␈α∞to
␈↓ α←␈↓reinterpret␈α∞by␈α∞making␈α∞the␈α∞smallest␈α∞change␈α∞it␈α∞can,␈α∞acting␈α∞on␈α∞the␈α∞assumption
␈↓ α←␈↓that␈α∞the␈α
original␈α∞clause␈α∞is␈α
probably␈α∞close␈α
to␈α∞correct.␈α∞ If␈α
this␈α∞fails,␈α∞the␈α
clause
␈↓ α←␈↓with the next highest overall score is chosen.
␈↓ α←␈↓␈↓ β?The␈α∞alternatives␈α∞are␈α∂presented␈α∞one␈α∞by␈α∞one␈α∂for␈α∞approval.␈α∞If␈α∂none␈α∞of
␈↓ α←␈↓them␈αis␈αacceptable,␈αthe␈αsystem␈αprompts␈αfor␈αa␈αnew␈αline␈αof␈αtext.␈α There␈αare␈αtwo
␈↓ α←␈↓reasons␈α∞for␈α∞showing␈α∞all␈α∞the␈α
current␈α∞alternatives␈α∞before␈α∞asking␈α∞for␈α∞new␈α
text.
␈↓ α←␈↓First,␈α∞the␈α∞selection␈α∞of␈α∞an␈α∂alternative␈α∞is␈α∞very␈α∞fast,␈α∞because␈α∞all␈α∂necessary␈α∞data
␈↓ α←␈↓structures␈α∞are␈α∂already␈α∞present.␈α∞ Reprocessing␈α∂a␈α∞new␈α∞line␈α∂of␈α∞text␈α∂would␈α∞take
␈↓ α←␈↓much␈αlonger.␈α More␈αimportant,␈αhowever,␈αthe␈αexpert␈αgets␈αa␈αgood␈αidea␈αof␈αwhat
␈↓ α←␈↓he␈α∞might␈α
have␈α∞said␈α
that␈α∞triggered␈α
the␈α∞system's␈α
incorrect␈α∞interpretations␈α
and
␈↓ α←␈↓can␈α
then␈α
rephrase␈αhis␈α
statement␈α
appropriately.␈α This␈α
can␈α
become␈αimportant␈α
if
␈↓ α←␈↓there␈α∩are␈α∩only␈α∩subtle␈α∪differences␈α∩in␈α∩the␈α∩translations␈α∩of␈α∪several␈α∩functions,
␈↓ α←␈↓attributes, etc.
␈↓ α←␈↓␈↓ β?This␈α∂segment␈α∂also␈α∂demonstrates␈α∂the␈α∞utility␈α∂of␈α∂even␈α∂the␈α∂very␈α∞limited
␈↓ α←␈↓form␈α
of␈α∞consistency␈α
checking␈α∞that␈α
is␈α
available.␈α∞ ␈↓¬TEIRESIAS␈↓␈α
maintains␈α∞a␈α
record
␈↓ α←␈↓of␈α⊃the␈α⊃tree␈α⊃of␈α⊃parses,␈α⊃and␈α∩whenever␈α⊃a␈α⊃clause␈α⊃is␈α⊃changed␈α⊃by␈α∩selecting␈α⊃an
␈↓ α←␈↓alternative,␈αit␈αchecks␈αto␈αmake␈αsure␈αthat␈αthe␈αnew␈αalternative␈αis␈αconsistent␈αwith
␈↓ α←␈↓all␈α∂the␈α∂other␈α∂clauses␈α⊂derived␈α∂from␈α∂the␈α∂same␈α⊂line␈α∂of␈α∂text.␈α∂ Once␈α⊂again,␈α∂the
␈↓ α←␈↓process is very fast, since all the data structures already exist.
␈↓ α←␈↓␈↓104    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬Anything else?
␈↓"β␈↓ α←␈↓¬++**␈↓π D 3␈↓¬
␈↓"β␈↓ α←␈↓¬clause(s) 3 deleted.

␈↓"β␈↓ α←␈↓¬Anything else?
␈↓"β␈↓ α←␈↓¬++**␈↓π ?␈↓¬
␈↓"β␈↓ α←␈↓¬     1 - the infection is primary-bacteremia
␈↓"β␈↓ α←␈↓¬ and 2 -   <deleted>
␈↓"β␈↓ α←␈↓¬ and 3 -   <deleted>
␈↓"β␈↓ α←␈↓¬ and 4 - the organism is not normally found at this site

␈↓"β␈↓ α←␈↓¬Anything else?
␈↓"β␈↓ α←␈↓¬++**␈↓π U 3␈↓¬
␈↓"β␈↓ α←␈↓¬clause(s) 3 undeleted.
␈↓"β␈↓ α←␈↓¬_______________________________________

␈↓ α←␈↓␈↓ β?No␈α
deletions␈αare␈α
actually␈αdone␈α
until␈αexiting␈α
from␈αthe␈α
editor,␈αmaking␈α
it
␈↓ α←␈↓possible to undo mistaken deletions.

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬Anything else?
␈↓"β␈↓ α←␈↓¬++**␈↓π D 4␈↓¬
␈↓"β␈↓ α←␈↓¬clause(s) 4 deleted.

␈↓"β␈↓ α←␈↓¬Anything else?
␈↓"β␈↓ α←␈↓¬++**␈↓π C 3␈↓¬
␈↓"β␈↓ α←␈↓¬Suppose (3) were changed to read-
␈↓"β␈↓ α←␈↓¬        3 - The site of the culture is one of: those sites that are normally sterile
␈↓"β␈↓ α←␈↓¬Would that be ok?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬

␈↓"β␈↓ α←␈↓¬Anything else?
␈↓"β␈↓ α←␈↓¬++**␈↓π ?␈↓¬
␈↓"β␈↓ α←␈↓¬     1 - the infection is primary-bacteremia
␈↓"β␈↓ α←␈↓¬ and 2 -   <deleted>
␈↓"β␈↓ α←␈↓¬ and 3 - the site of the culture is one of: those sites that are normally sterile
␈↓"β␈↓ α←␈↓¬ and 4 -   <deleted>

␈↓"β␈↓ α←␈↓¬++**␈↓π OK␈↓¬

␈↓"β␈↓ α←␈↓¬Change action?
␈↓"β␈↓ α←␈↓¬--**␈↓π N␈↓¬

␈↓"β␈↓ α←␈↓¬This is my understanding of your rule:
␈↓"β␈↓ α←␈↓¬RULE383
␈↓"β␈↓ α←␈↓¬-------
␈↓"β␈↓ α←␈↓¬  If:  1) The infection is primary-bacteremia, and
␈↓"β␈↓ α←␈↓¬       2) The site of the culture is one of: those sites that are normally sterile
␈↓"β␈↓ α←␈↓¬  Then:  There is strongly suggestive evidence (.8) that the category of the
␈↓"β␈↓ α←␈↓¬         organism is enterobacteriaceae

␈↓"β␈↓ α←␈↓¬Okay? (Yes or No)
␈↓"β␈↓ α←␈↓¬--**␈↓π Y
␈↓"β␈↓ α←␈↓π_______________________________________

␈↓ α←␈↓␈↓ β?After␈α
making␈α
all␈α
the␈α
changes␈αthat␈α
the␈α
expert␈α
wanted,␈α
the␈αsystem␈α
again
␈↓ α←␈↓displays␈α∂the␈α∂rule␈α∂and␈α∂asks␈α⊂for␈α∂approval.␈α∂ This␈α∂time␈α∂the␈α∂expert␈α⊂is␈α∂satisfied
␈↓ α←␈↓with the interpretation.
␈↓ α←␈↓␈↓5-4␈↓ λ⊃HOW IT ALL WORKS    105␈↓

␈↓"β␈↓ α←␈↓␈↓ β?This␈α∞segment␈α∞of␈α
the␈α∞trace␈α∞also␈α∞displays␈α
one␈α∞weakness␈α∞of␈α∞the␈α
current
␈↓ α←␈↓implementation: ␈α∞Efficient␈α∞use␈α∞of␈α∞the␈α∞rule␈α∞editor␈α∞requires␈α∞a␈α∞familiarity␈α
with
␈↓ α←␈↓the␈αnature␈αof␈αthe␈αmistakes␈αproduced␈αby␈αthe␈αnatural␈αlanguage␈αroutines.␈α Since
␈↓ α←␈↓all␈α↔four␈α⊗lines␈α↔of␈α↔the␈α⊗system's␈α↔original␈α⊗interpretation␈α↔are␈α↔incorrect,␈α⊗the
␈↓ α←␈↓command␈α∂``CHANGE␈α∞1␈α∂2␈α∞3␈α∂4''␈α∞seems␈α∂plausible.␈α∞ It␈α∂takes␈α∞a␈α∂certain␈α∂level␈α∞of
␈↓ α←␈↓familiarity␈α∀with␈α∀␈↓¬TEIRESIAS␈↓␈α∀to␈α∀realize␈α∃that␈α∀clauses␈α∀1␈α∀and␈α∀3␈α∀are␈α∃close␈α∀but
␈↓ α←␈↓incorrect,␈α⊂while␈α⊃2␈α⊂and␈α⊂4␈α⊃are␈α⊂purely␈α⊂spurious.␈α⊃ While␈α⊂the␈α⊂expert␈α⊃need␈α⊂not
␈↓ α←␈↓understand␈α
how␈α
the␈α
natural␈α
language␈α
routines␈α
work,␈α
and␈α
might␈α
acquire␈αthe
␈↓ α←␈↓necessary␈αsophistication␈αthrough␈αexperience,␈αthe␈αnontransparency␈αof␈αthis␈αpart
␈↓ α←␈↓of the system still presents a problem, and is a likely topic for future work.

␈↓ α←␈↓␈↓α5-4-4    Second guessing␈↓
␈↓ α←␈↓␈↓ ¬GDo␈α
you␈αknow␈α
what␈αyou␈α
are␈αdoing?␈α
 Will␈αyou␈α
listen
␈↓ α←␈↓␈↓ ¬Gto words to answer yours, and then pass judgment?
␈↓"β␈↓ α←␈↓␈↓ π+␈↓↓Oedipus the King␈↓, lines 543-544
␈↓ α←␈↓␈↓ β?Now␈α
that␈αthe␈α
expert␈α
has␈αindicated␈α
that␈α
the␈αinterpretation␈α
of␈α
his␈αtext
␈↓ α←␈↓is␈αcorrect,␈α␈↓¬TEIRESIAS␈↓␈αdouble-checks␈αthe␈αrule.␈α The␈αbasic␈αidea␈αis␈αto␈αuse␈αthe␈αrule
␈↓ α←␈↓model␈α⊂to␈α⊂see␈α∂how␈α⊂well␈α⊂this␈α∂new␈α⊂rule␈α⊂``fits␈α∂in''␈α⊂to␈α⊂the␈α∂system's␈α⊂model␈α⊂of␈α∂its
␈↓ α←␈↓knowledge, that is, does it ``look like'' a typical rule of the sort expected?
␈↓ α←␈↓␈↓ β?The␈αpoint␈αhere␈αis␈αto␈αtake␈αadvantage␈αof␈αseveral␈α
unique␈αcharacteristics
␈↓ α←␈↓of␈αthe␈α``student''␈α
being␈αtutored: ␈α␈↓¬TEIRESIAS␈↓␈α
has␈α``total␈αrecall''␈α
of␈αevery␈αrule␈αin␈α
the
␈↓ α←␈↓knowledge␈α
base␈α
and␈α
has␈α
a␈α
great␈α
capacity␈α
for␈α
dealing␈α
with␈α
large␈α
amounts␈α
of
␈↓ α←␈↓detail.␈α∞ Both␈α∞of␈α∞these␈α∞characteristics␈α∞are␈α∞put␈α∞to␈α∞use␈α∞in␈α∞constructing␈α∂the␈α∞rule
␈↓ α←␈↓models.␈α Since␈α
the␈αexpert␈αmay␈α
be␈αexpressing␈αrules␈α
that␈αhave␈αnever␈α
previously
␈↓ α←␈↓been␈α∞formalized,␈α∞any␈α∞help␈α
that␈α∞the␈α∞system␈α∞can␈α
offer␈α∞will␈α∞prove␈α∞very␈α
useful.
␈↓ α←␈↓The␈α
models␈α
help␈α
by␈α
providing␈α
a␈αbasis␈α
for␈α
suggesting␈α
details␈α
that␈α
may␈αhave
␈↓ α←␈↓been␈αoverlooked.␈α In␈αdoing␈αso,␈αthey␈αalso␈αgive␈αthe␈αexpert␈αa␈αhint␈αof␈αthe␈α``world
␈↓ α←␈↓view'' implicit in the rules already in the knowledge base.
␈↓ α←␈↓␈↓ β?The␈α∪presence␈α∀of␈α∪a␈α∪partial␈α∀match␈α∪between␈α∪the␈α∀new␈α∪rule␈α∀and␈α∪the
␈↓ α←␈↓generalizations␈αin␈α
the␈αrule␈α
model␈αtrigger␈α
a␈αresponse␈α
from␈αthe␈α
system.␈α Recall
␈↓ α←␈↓the last ntuple of the premise description in the rule model of Fig. 5-3:

␈↓"β␈↓ α←␈↓	␈↓ β↔((SITE MEMBF SAME) (INFECTION SAME) (PORTAL SAME) 1.23))

␈↓ α←␈↓This␈αmeans␈α
that␈αwhen␈α
culture␈α␈↓	SITE␈↓␈αappears␈α
in␈αthe␈α
rule,␈α␈↓	INFECTION␈↓␈αtype␈α
and
␈↓ α←␈↓organism␈α
␈↓	PORTAL␈↓␈α
often␈α
appear␈α
as␈α
well.␈α
 The␈α
new␈α
rule␈α
has␈α
the␈α
first␈α
two␈α
but
␈↓ α←␈↓lacks the last, and the system points this out:

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬I hate to criticize, Dr. Davis, but did you know that most rules about what the
␈↓"β␈↓ α←␈↓¬category of a organism might be, that mention-
␈↓"β␈↓ α←␈↓¬            the site of a culture
␈↓"β␈↓ α←␈↓¬            the infection
␈↓"β␈↓ α←␈↓¬ALSO mention-
␈↓"β␈↓ α←␈↓¬     [A] - the portal of entry of a organism

␈↓"β␈↓ α←␈↓¬Shall I try to write a clause to account for [A] ?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y␈↓¬
␈↓"β␈↓ α←␈↓¬How about-
␈↓"β␈↓ α←␈↓¬     [A] The portal of entry of the organism is gi
␈↓ α←␈↓␈↓106    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓¬Ok?
␈↓"β␈↓ α←␈↓¬++**␈↓π Y
␈↓"β␈↓ α←␈↓π_______________________________________

␈↓ α←␈↓␈↓ β?Since␈α⊂the␈α⊂expert␈α⊂has␈α⊂agreed␈α⊂to␈α⊂a␈α⊂clause␈α⊂about␈α⊂portal␈α⊂of␈α⊃entry,␈α⊂the
␈↓ α←␈↓system␈αconstructs␈αit.␈α The␈αpredicate␈αfunction␈α(␈↓	SAME␈↓)␈αis␈αsupplied␈αby␈αthe␈α
ntuple.
␈↓ α←␈↓The␈α∞template␈α
for␈α∞␈↓	SAME␈↓␈α∞is␈α
then␈α∞completed␈α
in␈α∞the␈α∞usual␈α
way,␈α∞except␈α∞that␈α
the
␈↓ α←␈↓system␈αfirst␈α
checks␈αin␈αthe␈α
record␈αof␈αthe␈α
consultation.␈α In␈αthis␈α
case,␈αit␈α
needs␈αa
␈↓ α←␈↓␈↓	VALUE␈↓␈α∂for␈α∂␈↓	PORTAL␈↓,␈α∂which␈α∂is␈α∂available␈α∞as␈α∂the␈α∂answer␈α∂to␈α∂question␈α∂18,␈α∞asked
␈↓ α←␈↓during␈α⊂the␈α⊂consultation␈α⊂(Section␈α⊂3-10).␈α⊂ Nothing␈α⊂further␈α⊂is␈α⊂needed,␈α⊃so␈α⊂the
␈↓ α←␈↓system␈α⊗requests␈α⊗no␈α⊗assistance.␈α↔ (Had␈α⊗the␈α⊗desired␈α⊗information␈α↔not␈α⊗been
␈↓ α←␈↓available--as␈α
would␈αbe␈α
the␈αcase␈α
if␈αit␈α
had␈α
``come␈αin␈α
cold''␈αrather␈α
than␈αfrom␈α
the
␈↓ α←␈↓consultation--the␈α∩expert␈α∩would␈α⊃have␈α∩been␈α∩asked␈α⊃to␈α∩supply␈α∩whatever␈α⊃was
␈↓ α←␈↓needed.)␈αThe␈α
result␈αis␈αa␈α
plausible␈αguess,␈αsince␈α
it␈αinsures␈αthat␈α
the␈αrule␈α
will␈αin
␈↓ α←␈↓fact␈α∪work␈α∪for␈α∪the␈α∪current␈α∪case␈α∪(note␈α∪the␈α∪further␈α∪use␈α∪of␈α∀the␈α∪``knowledge
␈↓ α←␈↓acquisition␈αin␈αcontext''␈αidea).␈α It␈αis␈αnot␈αnecessarily␈αcorrect,␈αof␈αcourse,␈αsince␈αthe
␈↓ α←␈↓desired clause may be more general, but it is at least a plausible attempt.

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬Also, most rules about what the category of a organism might be ALSO conclude about-
␈↓"β␈↓ α←␈↓¬     [A] - the identity of a organism
␈↓"β␈↓ α←␈↓¬Shall I try to write a clause to account for [A] ?
␈↓"β␈↓ α←␈↓¬++**␈↓π N
␈↓"β␈↓ α←␈↓π_______________________________________

␈↓ α←␈↓␈↓ β?There␈αare␈αalso␈αntuples␈α(in␈αthe␈αdescription␈αof␈αthe␈αaction)␈αthat␈αindicate
␈↓ α←␈↓patterns␈α
in␈α
the␈αaction␈α
of␈α
the␈αrule,␈α
and␈α
one␈αof␈α
these␈α
is␈αapplicable␈α
to␈α
the␈αnew
␈↓ α←␈↓rule.␈α
 As␈αalways,␈α
the␈αexpert␈α
can␈α
override␈αthe␈α
system's␈αsuggestions,␈α
and␈αdoes␈α
so
␈↓ α←␈↓in this case.
␈↓ α←␈↓␈↓ ¬GSo␈αclear␈αin␈αthis␈αcase␈αwere␈αthe␈αoracles;␈αso␈αclear␈αand
␈↓ α←␈↓␈↓ ¬Gyet false.
␈↓"β␈↓ α←␈↓␈↓ π+␈↓↓Oedipus the King␈↓, lines 722-723

␈↓ α←␈↓␈↓ β?It␈α∂should␈α∂be␈α∂noted␈α∂that␈α∂there␈α∂is␈α∂nothing␈α∂in␈α∂this␈α∂concept␈α∂of␈α∂``second
␈↓ α←␈↓guessing''␈αthat␈αis␈αspecific␈αto␈αthe␈αrule␈αmodels␈αas␈αthey␈αare␈αcurrently␈αdesigned,␈α
or
␈↓ α←␈↓indeed␈α∂to␈α∂the␈α⊂rules␈α∂as␈α∂a␈α∂knowledge␈α⊂representation.␈α∂ The␈α∂most␈α⊂general␈α∂and
␈↓ α←␈↓fundamental␈αpoint␈αis␈α
that␈αmentioned␈αabove,␈αof␈α
testing␈αto␈αsee␈α
how␈αsomething
␈↓ α←␈↓``fits␈α∂in''␈α∂to␈α⊂the␈α∂system's␈α∂model␈α⊂of␈α∂its␈α∂knowledge.␈α⊂ At␈α∂this␈α∂point,␈α⊂the␈α∂system
␈↓ α←␈↓might␈α
perform␈α
any␈αkind␈α
of␈α
check,␈α
for␈αviolations␈α
of␈α
any␈αestablished␈α
prejudices
␈↓ α←␈↓about␈α∞what␈α
the␈α∞new␈α
chunk␈α∞of␈α
knowledge␈α∞should␈α
look␈α∞like.␈α
 For␈α∞rules,␈α
these
␈↓ α←␈↓checks␈αmight␈αconcern␈αthe␈αsize␈αof␈αits␈αcertainty␈αfactor,␈αthe␈αnumber␈αof␈αclauses␈αin
␈↓ α←␈↓the␈αpremise,␈αetc.,␈αin␈αaddition␈αto␈αthe␈αcurrent␈αchecks.␈α Checks␈αused␈αwith,␈αsay,␈αa
␈↓ α←␈↓procedural␈α∀encoding␈α∃might␈α∀involve␈α∀the␈α∃number␈α∀and␈α∀type␈α∃of␈α∀arguments
␈↓ α←␈↓passed␈α⊂to␈α⊂the␈α⊂procedure,␈α⊂the␈α⊃use␈α⊂of␈α⊂global␈α⊂variables,␈α⊂the␈α⊂presence␈α⊃of␈α⊂side
␈↓ α←␈↓effects,␈α∃etc.␈α∃ In␈α∃that␈α∀case,␈α∃for␈α∃example,␈α∃we␈α∀can␈α∃imagine␈α∃adding␈α∃a␈α∀new
␈↓ α←␈↓procedure␈α⊂to␈α∂a␈α⊂system␈α⊂that␈α∂then␈α⊂responds␈α∂by␈α⊂remarking␈α⊂``...␈↓↓most␈α∂procedures
␈↓ α←␈↓↓that␈α∩do␈α∩hash-table␈α∪insertion␈α∩also␈α∩have␈α∩the␈α∪side␈α∩effect␈α∩of␈α∪incrementing␈α∩the
␈↓ α←␈↓↓variable␈α
␈↓	NUMBERELEMENTS.␈↓↓␈αShall␈α
I␈α
add␈αthe␈α
code␈α
to␈αdo␈α
this?''␈↓  In␈α
general,␈αthis
␈↓ α←␈↓␈↓5-4␈↓ λ⊃HOW IT ALL WORKS    107␈↓

␈↓"β␈↓ α←␈↓``second␈αguessing''␈αprocess␈αcan␈α
involve␈αany␈αcharacteristic␈αthat␈αthe␈α
system␈αmay
␈↓ α←␈↓have ``noticed'' about the particular knowledge representation in use.

␈↓ α←␈↓␈↓α5-4-5    Final checkout␈↓
␈↓ α←␈↓␈↓ β?Now␈α
that␈α
both␈αthe␈α
expert␈α
and␈α
␈↓¬TEIRESIAS␈↓␈αare␈α
satisfied,␈α
there␈α
is␈αone␈α
final
␈↓ α←␈↓sequence␈α∀of␈α∀tests␈α∀to␈α∀be␈α∀performed,␈α∀reflecting␈α∀once␈α∀again␈α∀the␈α∀benefit␈α∀of
␈↓ α←␈↓knowledge acquisition in context.
␈↓ α←␈↓␈↓ β?At␈α⊗this␈α⊗point,␈α∃␈↓¬TEIRESIAS␈↓␈α⊗examines␈α⊗several␈α∃things␈α⊗about␈α⊗the␈α∃rule,
␈↓ α←␈↓attempting␈α
to␈α
make␈α
sure␈α
that␈α
the␈αrule␈α
will,␈α
in␈α
fact,␈α
fix␈α
the␈αproblem␈α
uncovered.
␈↓ α←␈↓In␈α
this␈α
case,␈α
for␈α
instance,␈α
the␈α
action␈α
of␈α
the␈α
new␈α
rule␈α
should␈α
be␈α
a␈αconclusion
␈↓ α←␈↓about␈αcategory;␈αthe␈αcategory␈αmentioned␈αshould␈αbe␈αenterobacteriaceae,␈αand␈αthe
␈↓ α←␈↓conclusion␈α
should␈α
be␈α
affirmative.␈α
The␈α
premise␈α
should␈α
not␈α
contain␈αany␈α
clauses
␈↓ α←␈↓that␈αare␈αsure␈αto␈αfail␈αin␈αthe␈αcontext␈αin␈αwhich␈αthe␈αrule␈αwill␈αbe␈αinvoked.␈α (Note
␈↓ α←␈↓that␈α
these␈α∞tests␈α
require␈α∞the␈α
ability␈α
to␈α∞dissect␈α
and␈α∞partially␈α
evaluate␈α∞the␈α
rule,
␈↓ α←␈↓and␈α⊂they␈α⊃rely␈α⊂heavily␈α⊂on␈α⊃the␈α⊂use␈α⊃of␈α⊂the␈α⊂function␈α⊃templates␈α⊂to␈α⊃guide␈α⊂that
␈↓ α←␈↓dissection.)␈α∪All␈α∪of␈α∪these␈α∪are␈α∪potential␈α∪sources␈α∪of␈α∪error␈α∪that␈α∀would␈α∪make
␈↓ α←␈↓obvious the fact that the rule will not fix the bug.
␈↓ α←␈↓␈↓ β?The␈α∞new␈α
rule␈α∞in␈α
the␈α∞current␈α∞example␈α
passes␈α∞all␈α
these␈α∞tests,␈α∞but␈α
note
␈↓ α←␈↓what␈α
would␈α
have␈α
happened␈α
if␈αthe␈α
user␈α
had␈α
(accidentally)␈α
typed␈α
initially␈α␈↓↓the
␈↓ α←␈↓↓site␈αis␈α␈↓αnot␈↓↓␈αone␈αof␈αthe␈αsterile␈αsites␈↓.␈α Everything␈αwould␈αhave␈αbeen␈αthe␈αsame␈αup
␈↓ α←␈↓until this point, when the following exchange would have occurred.

␈↓"β␈↓ α←␈↓	␈↓π________________________________________␈↓	
␈↓"β␈↓ α←␈↓	The rule now reads:
␈↓"β␈↓ α←␈↓	RULE383
␈↓"β␈↓ α←␈↓	-------
␈↓"β␈↓ α←␈↓	  If:  1) The infection is primary-bacteremia, and
␈↓"β␈↓ α←␈↓	       2) The site of the culture is not one of: those sites
␈↓"β␈↓ α←␈↓	          that are normally sterile, and
␈↓"β␈↓ α←␈↓	       3) The portal of entry of the organism is gi
␈↓"β␈↓ α←␈↓	  Then:  There is strongly suggestive evidence (.8) that the
␈↓"β␈↓ α←␈↓	         category of the organism is enterobacteriaceae

␈↓"β␈↓ α←␈↓	Sorry, but this rule won␈↓'␈↓	t work.  In the context of
␈↓"β␈↓ α←␈↓	ORGANISM-1, clause 2 ("the site of the culture is not one of:
␈↓"β␈↓ α←␈↓	those sites that are normally sterile") is sure to fail.

␈↓"β␈↓ α←␈↓	Would you like to change the rule?
␈↓"β␈↓ α←␈↓	++**
␈↓"β␈↓ α←␈↓	␈↓π________________________________________

␈↓ α←␈↓␈↓ β?The␈α
expert␈α
then␈α
has␈α
the␈α
option␈α
of␈α
either␈α
editing␈α
the␈α
current␈α
rule␈α
or
␈↓ α←␈↓writing␈α∩an␈α∩entirely␈α⊃new␈α∩one␈α∩(since␈α⊃the␈α∩current␈α∩rule␈α⊃may␈α∩be␈α∩correct,␈α⊃only
␈↓ α←␈↓inapplicable␈α
to␈α
the␈α
current␈α∞problem).␈α
 If␈α
he␈α
edits␈α∞it,␈α
the␈α
tests␈α
are␈α∞run␈α
again,
␈↓ α←␈↓until␈α∞the␈α∞system␈α∞is␈α
satisfied␈α∞that␈α∞there␈α∞is␈α
nothing␈α∞obviously␈α∞wrong␈α∞with␈α
the
␈↓ α←␈↓rule.
␈↓ α←␈↓␈↓108    KNOWLEDGE ACQUISITION I␈↓ 
#5-4␈↓

␈↓"β␈↓ α←␈↓␈↓α5-4-6    Bookkeeping␈↓
␈↓ α←␈↓␈↓ β?There␈α∪are␈α∪a␈α∪number␈α∪of␈α∪straightforward␈α∪bookkeeping␈α∪tasks␈α∀to␈α∪be
␈↓ α←␈↓performed.␈α Some␈αof␈αthem␈αinvolve␈αhooking␈αthe␈αnew␈αrule␈αinto␈αthe␈αknowledge
␈↓ α←␈↓base␈αso␈αthat␈αit␈αis␈αretrieved␈αand␈αinvoked␈αappropriately.␈α ␈↓¬TEIRESIAS␈↓␈αdoes␈αthis␈αby
␈↓ α←␈↓scanning␈α∞the␈α∞clauses␈α∞of␈α
the␈α∞action␈α∞part␈α∞and␈α
adding␈α∞the␈α∞rule␈α∞number␈α∞to␈α
the
␈↓ α←␈↓proper␈α
internal␈α
lists␈α
(e.g.,␈α
in␈α
this␈α
case,␈α
it␈α
adds␈α
the␈α
rule␈α
number␈α
to␈α
the␈α
list␈αof
␈↓ α←␈↓rules␈αthat␈αconclude␈αabout␈α␈↓	CATEGORY␈↓.)␈αThis␈αtask␈αis␈αnot␈αdifficult␈αbecause␈α
meta-
␈↓ α←␈↓rules␈α∂can␈α∂be␈α∂used␈α∂to␈α∂express␈α∂all␈α∂necessary␈α∂ordering␈α∂of␈α∂rule␈α⊂invocation␈α∂(see
␈↓ α←␈↓chapter␈α
7).␈α
 As␈α∞a␈α
result,␈α
the␈α∞lists␈α
can␈α
be␈α
stored␈α∞unordered,␈α
so␈α
adding␈α∞a␈α
new
␈↓ α←␈↓rule␈αto␈αthem␈αis␈αeasy.␈α All␈αof␈αthis␈αis␈αtaken␈αcare␈αof␈αwithout␈αassistance␈αfrom␈αthe
␈↓ α←␈↓user.
␈↓ α←␈↓␈↓ β?Another␈α∩of␈α∩the␈α∩bookkeeping␈α∪tasks␈α∩involves␈α∩tagging␈α∩the␈α∪rule␈α∩with
␈↓ α←␈↓information␈α
which␈α
facilitates␈α
maintaining␈α
the␈α
large␈α
and␈α
constantly␈αchanging
␈↓ α←␈↓body␈αof␈αrules.␈α Each␈αrule␈αis␈αtagged␈αwith␈αthe␈αname␈αof␈αthe␈αauthor,␈αthe␈αdate␈αof
␈↓ α←␈↓creation,␈αthe␈αcase␈αnumber␈α
that␈αprompted␈αits␈αcreation␈α
(if␈αthere␈αis␈αone),␈αand␈α
the
␈↓ α←␈↓user's␈αown␈αcomments␈αon␈αwhy␈αhe␈αadded␈α(or␈αchanged)␈αthe␈αrule.␈α This␈αlast␈αitem
␈↓ α←␈↓gives␈α
the␈α
expert␈α
a␈α
way␈α
of␈α
recording␈αthe␈α
reasons␈α
for␈α
any␈α
rules␈α
or␈α
parts␈αof␈α
rules
␈↓ α←␈↓that␈α
are␈αnot␈α
obvious.␈α Since␈α
the␈α
rule␈αin␈α
this␈αexample␈α
is␈αfairly␈α
straightforward,
␈↓ α←␈↓so are the user's comments.

␈↓"β␈↓ α←␈↓¬_______________________________________
␈↓"β␈↓ α←␈↓¬Please describe briefly why you added (or changed) this rule.
␈↓"β␈↓ α←␈↓¬[Type an empty line when done.]
␈↓"β␈↓ α←␈↓¬**␈↓πTHE SYSTEM FAILED TO REALIZE IT COULD CONCLUDE CATEGORY, AND THIS ALLOWED RULE184␈↓¬
␈↓"β␈↓ α←␈↓¬**␈↓πTO INCORRECTLY CONCLUDE IDENTITY.␈↓¬
␈↓"β␈↓ α←␈↓¬**

␈↓"β␈↓ α←␈↓¬RULE383 has now been added to the knowledge base.
␈↓"β␈↓ α←␈↓¬_______________________________________

␈↓ α←␈↓␈↓ β?It␈α∪should␈α∩be␈α∪noted␈α∩that␈α∪rule␈α∩383␈α∪gets␈α∩added␈α∪only␈α∩to␈α∪the␈α∩current
␈↓ α←␈↓``working''␈α∞knowledge␈α
base.␈α∞ All␈α
changes␈α∞made␈α
by␈α∞any␈α
individual␈α∞are␈α
stored
␈↓ α←␈↓away␈αat␈αthe␈αend␈αof␈αthe␈αsession,␈αfiled␈αunder␈αthe␈αindividual's␈αname.␈α When␈αthe
␈↓ α←␈↓expert␈α⊃signs␈α⊂on␈α⊃again,␈α⊂the␈α⊃system␈α⊂automatically␈α⊃searches␈α⊂for␈α⊃a␈α⊂file␈α⊃of␈α⊂the
␈↓ α←␈↓proper␈αname␈αand␈αasks␈αif␈αthe␈αchanges␈αshould␈αbe␈αreinstated.␈α This␈αallows␈αeach
␈↓ α←␈↓expert␈α⊗to␈α⊗build␈α⊗up␈α⊗a␈α⊗knowledge␈α⊗base␈α⊗that␈α⊗includes␈α⊗his␈α⊗own␈α∃personal
␈↓ α←␈↓preferences␈α∂and␈α∂yet␈α∂does␈α∂not␈α∂cause␈α∂problems␈α∂with␈α∂maintaining␈α∂a␈α∂standard,
␈↓ α←␈↓functioning␈α∞system␈α∞for␈α∞other␈α∞users.␈α∞ Permanent␈α∞changes␈α∞to␈α∂the␈α∞performance
␈↓ α←␈↓program knowledge base are made after agreement from the experts.
␈↓ α←␈↓␈↓ β?At␈α∞this␈α∞point,␈α∞the␈α∞system␈α∞also␈α∞performs␈α∞any␈α∞necessary␈α
recomputation
␈↓ α←␈↓of␈αrule␈αmodels.␈↓
15␈↓␈αThe␈αoperation␈αis␈αvery␈αfast,␈αsince␈αit␈αis␈αclear␈αfrom␈α
the␈αaction
␈↓ α←␈↓part␈α
of␈α
the␈αrule␈α
which␈α
models␈αmay␈α
need␈α
to␈αbe␈α
recomputed,␈α
and␈αthe␈α
␈↓	EXAMPLES␈↓
␈↓ α←␈↓part␈α∂supplies␈α∂the␈α∂names␈α∂of␈α∂the␈α∂other␈α∂relevant␈α∂rules.␈α∂ Note␈α∂that␈α∂this␈α∞means

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[15]␈αThe␈αmodels␈αare␈αrecomputed␈αwhen␈αany␈αchange␈αis␈αmade␈αto␈αthe␈αknowledge
␈↓ α←␈↓base, including rule deletion or modification, as well as addition.
␈↓ α←␈↓␈↓5-4␈↓ λ⊃HOW IT ALL WORKS    109␈↓

␈↓"β␈↓ α←␈↓that␈α∂␈↓¬TEIRESIAS␈↓'s␈α∂model␈α∂of␈α∂the␈α∂knowledge␈α∂base␈α∂is␈α∂kept␈α∂constantly␈α∂up-to-date,
␈↓ α←␈↓immediately␈α_reflecting␈α↔any␈α_changes.␈α↔ As␈α_a␈α↔result,␈α_performance␈α_on␈α↔the
␈↓ α←␈↓acquisition␈αof␈α
subsequent␈αrules␈α
may␈αconceivably␈α
be␈αbetter.␈α
 Since␈αthe␈α
updated
␈↓ α←␈↓rule␈α
models␈α
are␈α
filed␈α
away␈α
with␈α
the␈α
other␈α
changes,␈α
this␈α
also␈α
means␈α
that␈α
the
␈↓ α←␈↓system␈α∂will␈α∂again␈α∂reflect␈α∂the␈α∂``structure''␈α∂of␈α∂this␈α∂expert's␈α∂reasoning␈α⊂the␈α∂next
␈↓ α←␈↓time he logs in.

␈↓ α←␈↓␈↓α5-4-7    Rerunning the consultation␈↓
␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓␈α∪then␈α∀invokes␈α∪the␈α∪performance␈α∀program␈α∪as␈α∀a␈α∪subprocess,
␈↓ α←␈↓rerunning␈α∞the␈α∞consultation␈α∞to␈α∞insure␈α∞that␈α∞additional␈α∞effects␈α∞of␈α∞the␈α∞new␈α
rule
␈↓ α←␈↓are␈α⊗discovered.␈α⊗ To␈α⊗do␈α⊗this,␈α⊗␈↓¬TEIRESIAS␈↓␈α⊗first␈α⊗erases␈α⊗everything␈α⊗from␈α∃the
␈↓ α←␈↓database␈αexcept␈αthe␈αexpert's␈αresponses␈αto␈αquestions␈αasked␈αduring␈αthe␈αoriginal
␈↓ α←␈↓consultation.␈α∩ When␈α∩the␈α∪consultation␈α∩is␈α∩rerun,␈α∪the␈α∩information-requesting
␈↓ α←␈↓routines look in the database for answers before asking the expert.
␈↓ α←␈↓␈↓ β?In this case, the single new rule has repaired all the problems.
␈↓ α←␈↓␈↓110    KNOWLEDGE ACQUISITION I␈↓ 
#5-5␈↓

␈↓"β␈↓ α←␈↓␈↓α5-5    OTHER USES FOR THE RULE MODELS␈↓
␈↓ α←␈↓␈↓ β?There␈α_are␈α_other␈α↔applications␈α_of␈α_the␈α↔rule␈α_models␈α_that␈α_help␈α↔to
␈↓ α←␈↓characterize␈α
their␈α
role␈α∞in␈α
allowing␈α
the␈α
system␈α∞to␈α
``know␈α
what␈α
it␈α∞knows,''␈α
and
␈↓ α←␈↓which␈α
make␈α∞plausible␈α
the␈α
claim␈α∞that␈α
they␈α
indicate␈α∞useful␈α
regularities␈α∞in␈α
the
␈↓ α←␈↓knowledge base.

␈↓ α←␈↓␈↓α5-5-1    ``Knowing what you know'':  Rule models as abstract descriptions
␈↓ α←␈↓α␈↓ βKof knowledge␈↓
␈↓ α←␈↓␈↓ β?The␈α↔rule␈α↔models␈α↔have␈α↔also␈α↔been␈α↔integrated␈α↔into␈α↔the␈α↔question-
␈↓ α←␈↓answering␈α
program␈α
that␈α
is␈α
part␈α
of␈α
␈↓¬MYCIN␈↓.␈α
 Previously,␈α
␈↓¬MYCIN␈↓␈α
would␈α
respond␈αto␈α
a
␈↓ α←␈↓question␈α
like␈α
␈↓↓How␈αdo␈α
you␈α
determine␈αthe␈α
identity␈α
of␈αan␈α
organism?␈↓␈α
by␈αtyping␈α
out
␈↓ α←␈↓the␈αnames␈αof␈αall␈αthe␈αrelevant␈αrules␈αand␈αasking␈αwhich␈αthe␈αuser␈αwanted␈αto␈αsee.
␈↓ α←␈↓But␈αa␈αrule␈αmodel,␈αas␈αa␈α
generalization␈αof␈αan␈αentire␈αclass␈αof␈αrules,␈α
answers␈αthe
␈↓ α←␈↓question too.

␈↓"β␈↓ α←␈↓	________________________________________

␈↓"β␈↓ α←␈↓	** ␈↓αHOW DO YOU DECIDE THAT AN ORGANISM IS␈↓	
␈↓"β␈↓ α←␈↓	   ␈↓αPSEUDOMONAS AERUGINOSA?␈↓	

␈↓"β␈↓ α←␈↓	Rules which conclude that the identity of the organism is
␈↓"β␈↓ α←␈↓	pseudomonas-aeruginosa generally use one or more of the
␈↓"β␈↓ α←␈↓	following pieces of information:
␈↓"β␈↓ α←␈↓	     the site of the culture
␈↓"β␈↓ α←␈↓	     the gram stain of the organism
␈↓"β␈↓ α←␈↓	     the morphology of the organism
␈↓"β␈↓ α←␈↓	Furthermore, the following relationships hold:
␈↓"β␈↓ α←␈↓	     The gram stain of the organism and the morphology of the
␈↓"β␈↓ α←␈↓	     organism tend to appear together in these rules.

␈↓"β␈↓ α←␈↓	RULE184, RULE116, RULE047, RULE085, and RULE040 conclude that
␈↓"β␈↓ α←␈↓	     the identity of the organism is pseudomonas-aeruginosa.
␈↓"β␈↓ α←␈↓	Which of these do you wish to see?
␈↓"β␈↓ α←␈↓	**
␈↓"β␈↓ α←␈↓	________________________________________

␈↓ α←␈↓By␈α
simply␈α
``reading''␈α
the␈α
rule␈α
model␈α
to␈α
the␈α
user,␈α
the␈α
system␈α
first␈α∞supplies␈α
an
␈↓ α←␈↓overview␈αof␈αthe␈αknowledge␈αin␈αthe␈α
relevant␈αrules␈αand␈αthen␈αallows␈αthe␈α
user␈αto
␈↓ α←␈↓examine␈α∞any␈α∞of␈α
those␈α∞rules␈α∞for␈α∞specific␈α
details.␈α∞ This␈α∞overview␈α∞suggests␈α
the
␈↓ α←␈↓structure␈αof␈αglobal␈αtrends␈αin␈αthe␈αknowledge␈αof␈αthe␈αexperts␈αwho␈αassembled␈α
the
␈↓ α←␈↓knowledge␈α
base␈α
and␈αthus␈α
helps␈α
to␈αdefine␈α
the␈α
overall␈αapproach␈α
to␈α
any␈αgiven
␈↓ α←␈↓topic.

␈↓ α←␈↓␈↓α5-5-2    ``Knowing what you don't know''␈↓
␈↓ α←␈↓␈↓ β?There␈α∂are␈α∂models␈α∂in␈α∂the␈α⊂current␈α∂system␈α∂made␈α∂from␈α∂between␈α⊂2␈α∂(the
␈↓ α←␈↓defined␈α
minimum)␈α
and␈α
35␈α
rules.␈α
 We␈α
have␈α
defined␈α
a␈α
metric␈α
to␈α
measure␈αthe
␈↓ α←␈↓``strength''␈α
of␈α∞a␈α
model␈α
and␈α∞have␈α
based␈α
it␈α∞on␈α
both␈α
the␈α∞total␈α
number␈α∞of␈α
rules
␈↓ α←␈↓from␈α∞which␈α∞the␈α∞model␈α∞was␈α∞constructed␈α∞and␈α∞the␈α∞size␈α∞of␈α∞the␈α∞CF␈α∞for␈α∂each␈α∞of
␈↓"β␈↓ α←␈↓␈↓5-5␈↓ εGOTHER USES FOR THE RULE MODELS    111␈↓

␈↓"β␈↓ α←␈↓those␈αrules.␈↓
16␈↓␈αThe␈αentire␈αmodel␈αset␈αis␈αkept␈αordered␈αfrom␈αweakest␈αto␈αstrongest,
␈↓ α←␈↓giving␈α⊃the␈α⊂system␈α⊃some␈α⊂indication␈α⊃of␈α⊃its␈α⊂likely␈α⊃competence␈α⊂on␈α⊃a␈α⊃range␈α⊂of
␈↓ α←␈↓subjects.␈α
 In␈α
a␈α∞very␈α
primitive␈α
way,␈α∞it␈α
thus␈α
gains␈α
a␈α∞knowledge␈α
of␈α
where␈α∞it␈α
is
␈↓ α←␈↓ignorant.
␈↓ α←␈↓␈↓ β?This␈α∞feature␈α
makes␈α∞possible␈α
a␈α∞new␈α
option␈α∞for␈α
the␈α∞expert.␈α∞Instead␈α
of
␈↓ α←␈↓adding␈αa␈αrule␈αto␈α
fix␈αa␈αbug,␈αhe␈α
can␈αask␈αthe␈αsystem␈α
what␈αit␈αwould␈αlike␈αto␈α
learn.
␈↓ α←␈↓(There␈αis␈α
currently␈αno␈α
real␈αlanguage␈α
processing␈αdone␈α
on␈αthe␈α
question␈αbelow;
␈↓ α←␈↓the system checks for precisely the sequence of words shown.)

␈↓"β␈↓ α←␈↓	________________________________________
␈↓"β␈↓ α←␈↓	--** ␈↓αWHAT WOULD YOU LIKE TO KNOW?␈↓	

␈↓"β␈↓ α←␈↓	There are only 2 rules for concluding that the identity of a
␈↓"β␈↓ α←␈↓	organism might be enterococcus. Would you like to give me a
␈↓"β␈↓ α←␈↓	rule for that?
␈↓"β␈↓ α←␈↓	++** ␈↓αN␈↓	

␈↓"β␈↓ α←␈↓	There are only 2 rules for concluding that a infection might be
␈↓"β␈↓ α←␈↓	infective-endocarditis.  Would you like to offer a rule for
␈↓"β␈↓ α←␈↓	that?
␈↓"β␈↓ α←␈↓	++** ␈↓αY␈↓	

␈↓"β␈↓ α←␈↓	The new rule will be called RULE384
␈↓"β␈↓ α←␈↓	 If     1-
␈↓"β␈↓ α←␈↓	________________________________________

␈↓ α←␈↓The␈α∞system␈α∞cycles␈α∞through␈α∞the␈α
rule␈α∞models␈α∞in␈α∞order,␈α∞indicating␈α∞the␈α
weakest
␈↓ α←␈↓topics␈α⊂first.␈α⊂ The␈α⊂current␈α⊂implementation␈α⊂ignores␈α⊂topics␈α⊂for␈α⊂which␈α⊂no␈α∂rule
␈↓ α←␈↓models␈α∩exist.␈α∪ It␈α∩also␈α∪makes␈α∩the␈α∩(imperfect)␈α∪assumption␈α∩that␈α∪subjects␈α∩for
␈↓ α←␈↓which␈α∪there␈α∪are␈α∪no␈α∪rules␈α∪at␈α∪all␈α∩are␈α∪the␈α∪sorts␈α∪of␈α∪things␈α∪that␈α∪would␈α∩not
␈↓ α←␈↓ordinarily be deduced.
␈↓ α←␈↓␈↓ β?This␈αis␈α
a␈αfirst␈α
order␈αsolution␈α
to␈αthe␈α
problem␈αof␈α
giving␈αthe␈α
system␈αan
␈↓ α←␈↓indication␈αof␈αits␈αweaknesses.␈α A␈αbetter␈αsolution␈αwould␈αsupply␈αan␈αindication␈α
of
␈↓ α←␈↓how␈αmuch␈αit␈αknows␈αabout␈αa␈αsubject,␈αas␈αcompared␈αwith␈αhow␈αmuch␈αthere␈αis␈αto
␈↓ α←␈↓know.␈α∞ There␈α∞surely␈α
are␈α∞subjects␈α∞for␈α
which␈α∞three␈α∞or␈α
four␈α∞rules␈α∞exhaust␈α
the
␈↓ α←␈↓available␈α∩knowledge,␈α⊃while␈α∩for␈α⊃others␈α∩a␈α⊃hundred␈α∩or␈α⊃more␈α∩rules␈α∩may␈α⊃not
␈↓ α←␈↓suffice.␈α The␈α
issue␈αis␈α
related␈αto␈α
work␈αdescribed␈α
in␈α[Carbonell73],␈α
on␈αclosed␈α
vs.
␈↓ α←␈↓open␈α∀sets.␈α∀ That␈α∀paper␈α∀offers␈α∀some␈α∀interesting␈α∀strategies␈α∀for␈α∀allowing␈α∪a
␈↓ α←␈↓program␈α
to␈αdecide␈α
when␈α
it␈αis␈α
ignorant␈α
and␈αhow␈α
it␈α
might␈αreason␈α
in␈α
the␈αface␈α
of
␈↓ α←␈↓the inability to store every fact about a given topic.
␈↓ α←␈↓␈↓ β?There␈αappear␈αto␈αbe␈αno␈αeasy␈αways␈αto␈αdeduce␈αthe␈αincompleteness␈αof␈αthe
␈↓ α←␈↓knowledge␈αbase␈α
using␈αonly␈α
the␈αinformation␈α
stored␈αin␈α
it.␈α It␈α
is␈αnot␈α
valid␈αto␈α
say,

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[16]␈α
In␈α
more␈α∞detail: ␈α
Model␈α
M1␈α∞is␈α
``stronger''␈α
than␈α∞model␈α
M2␈α
(a) if␈α∞M1␈α
was
␈↓ α←␈↓constructed␈α∂from␈α∞more␈α∂rules␈α∞than␈α∂M2,␈α∞or␈α∂(b) when␈α∞the␈α∂number␈α∞of␈α∂rules␈α∞in
␈↓ α←␈↓both␈αis␈αthe␈αsame,␈αif␈αthe␈αsum␈αof␈αthe␈αcertainty␈αfactors␈αof␈αthe␈αrules␈αfrom␈αwhich
␈↓ α←␈↓M1 was made is larger than the corresponding sum for M2.
␈↓"β␈↓ α←␈↓␈↓112    KNOWLEDGE ACQUISITION I␈↓ 
#5-5␈↓

␈↓"β␈↓ α←␈↓for␈α⊂instance,␈α∂that␈α⊂there␈α∂ought␈α⊂to␈α∂be␈α⊂even␈α∂a␈α⊂single␈α∂rule␈α⊂for␈α⊂every␈α∂attribute
␈↓ α←␈↓(how␈αcould␈αa␈αpatient's␈αname␈αbe␈αdeduced?).␈α Nor␈αis␈αthere␈αa␈αwell-defined␈αset␈αof
␈↓ α←␈↓attributes␈αfor␈α
which␈αno␈α
rules␈αare␈αlikely␈α
to␈αexist.␈α
 Nor␈αis␈αit␈α
clear␈αwhat␈α
sort␈αof
␈↓ α←␈↓information would allow the incompleteness to be deduced.
␈↓ α←␈↓␈↓ β?The␈α∞issue␈α∞is␈α∞a␈α∞significant␈α∞one,␈α∞since␈α∞a␈α∞good␈α∞solution␈α∞to␈α∞the␈α
problem
␈↓ α←␈↓would␈α∪not␈α∪only␈α∪give␈α∪␈↓¬TEIRESIAS␈↓␈α∀a␈α∪better␈α∪grasp␈α∪of␈α∪where␈α∀the␈α∪performance
␈↓ α←␈↓program␈α
was␈α
weak␈α
but␈α∞would␈α
also␈α
provide␈α
several␈α
important␈α∞capabilities␈α
to
␈↓ α←␈↓the␈α∃performance␈α∃program␈α∃itself.␈α∃ While␈α∃␈↓¬MYCIN␈↓␈α∃can␈α∃currently␈α∃recognize␈α∃a
␈↓ α←␈↓situation␈αin␈α
which␈αit␈α
is␈αunable␈α
to␈αmake␈α
recommendations␈αbecause␈α
of␈αa␈αlack␈α
of
␈↓ α←␈↓information,␈α⊃the␈α∩issue␈α⊃here␈α⊃is␈α∩␈↓↓why␈↓␈α⊃there␈α⊃is␈α∩too␈α⊃little␈α⊃information.␈α∩ If,␈α⊃for
␈↓ α←␈↓instance,␈α
it␈α
is␈α
unable␈α
to␈α
deduce␈α
a␈α
value␈α
for␈α
an␈α
attribute,␈α
is␈α
that␈α
because␈αthe
␈↓ α←␈↓expert␈α≤has␈α≤supplied␈α≠too␈α≤little␈α≤data␈α≠(answering␈α≤most␈α≤questions␈α≠with
␈↓ α←␈↓``unknown''),␈α⊂or␈α⊂is␈α⊃it␈α⊂because␈α⊂the␈α⊂program's␈α⊃knowledge␈α⊂of␈α⊂that␈α⊃attribute␈α⊂is
␈↓ α←␈↓incomplete?␈↓
17␈↓␈α
If␈α
the␈α
performance␈α
program␈αhad␈α
a␈α
way␈α
of␈α
judging␈α
the␈αextent
␈↓ α←␈↓of␈α∪its␈α∀ignorance,␈α∪it␈α∀would␈α∪be␈α∀able␈α∪to␈α∀distinguish␈α∪insufficient␈α∀data␈α∪from
␈↓ α←␈↓insufficient knowledge and take the proper action.
␈↓ α←␈↓␈↓ β?It␈α∂would␈α∂also␈α∞permit␈α∂the␈α∂use␈α∂of␈α∞the␈α∂``if␈α∂it␈α∞were␈α∂true␈α∂I␈α∂would␈α∞know''
␈↓ α←␈↓heuristic␈αin␈α
[Carbonell73].␈α Roughly␈α
restated,␈αthis␈αsays␈α
that␈α``if␈α
I␈αknow␈αa␈α
great
␈↓ α←␈↓deal␈αabout␈αsubject␈αS,␈αand␈αfact␈αF␈αconcerns␈αan␈αimportant␈αaspect␈αof␈αS,␈αthen␈αif␈αI
␈↓ α←␈↓don't␈α∩already␈α⊃know␈α∩that␈α⊃F␈α∩is␈α∩true,␈α⊃it's␈α∩probably␈α⊃false.'' ␈α∩Thus,␈α∩in␈α⊃certain
␈↓ α←␈↓circumstances␈α∀a␈α∃lack␈α∀of␈α∀knowledge␈α∃about␈α∀the␈α∀truth␈α∃of␈α∀a␈α∃statement␈α∀can
␈↓ α←␈↓plausibly be used as evidence suggesting that the statement is false.

␈↓ α←␈↓␈↓α5-6    PERSONALIZED WORLD VIEWS␈↓
␈↓ α←␈↓␈↓ β?One␈α∪of␈α∀the␈α∪likely␈α∀problems␈α∪of␈α∀collecting␈α∪knowledge␈α∀from␈α∪several
␈↓ α←␈↓experts␈α↔is␈α↔that␈α↔of␈α↔conflicting␈α↔world␈α↔views.␈α↔Since,␈α↔currently,␈α⊗individual
␈↓ α←␈↓modifications␈αto␈αthe␈αrule␈αbase␈αare␈αstored␈αaway␈αseparately␈αfor␈αeach␈αindividual
␈↓ α←␈↓user,␈αthe␈αestablished␈αknowledge␈αbase␈αis␈αkept␈αdistinct␈αfrom␈αlocal␈αmodifications.
␈↓ α←␈↓At␈α∪some␈α∪time,␈α∀however,␈α∪it␈α∪might␈α∀prove␈α∪useful␈α∪to␈α∀be␈α∪able␈α∪to␈α∀deal␈α∪with
␈↓ α←␈↓contributions␈αfrom␈α
several␈αexperts␈αand␈α
keep␈αthem␈αin␈α
the␈αknowledge␈α
base␈αall
␈↓ α←␈↓at␈αonce.␈α
 Some␈αvery␈α
limited␈αcapabilities␈α
in␈αthis␈α
direction␈αare␈α
made␈αpossible␈α
by
␈↓ α←␈↓tagging␈α
each␈αrule␈α
with␈α
the␈αname␈α
of␈αits␈α
author␈α
(as␈αillustrated␈α
earlier)␈α
and␈αby
␈↓ α←␈↓building a set of rule models for each individual's personal knowledge base.
␈↓ α←␈↓␈↓ β?Individualized␈α∞rule␈α∞models␈α∞make␈α∞it␈α∞possible,␈α∞for␈α∞instance,␈α∞to␈α∞ask␈α
the
␈↓ α←␈↓system␈α⊃␈↓↓How␈α⊃would␈α⊃Dr.␈α⊃Jones␈α⊃deduce␈α⊃the␈α⊃identity␈α⊃of␈α⊃an␈α⊃organism?␈↓␈α∩and␈α⊃to
␈↓ α←␈↓compare␈α∂this␈α∂with␈α∞the␈α∂reasoning␈α∂indicated␈α∞by␈α∂a␈α∂rule␈α∞model␈α∂for␈α∂a␈α∞different
␈↓ α←␈↓expert.␈α⊃ The␈α⊃tags␈α⊃on␈α⊃rules␈α⊃make␈α⊃it␈α⊃possible␈α⊃to␈α⊃focus␈α⊃conveniently␈α⊃on␈α⊃the
␈↓ α←␈↓effects␈α
of␈αthe␈α
contributions␈α
of␈αdifferent␈α
experts.␈α
We␈αcan␈α
imagine,␈αfor␈α
instance,
␈↓ α←␈↓a␈α
meta-rule␈α
of␈α
the␈α
form␈α
␈↓↓When␈α∞trying␈α
to␈α
deduce␈α
the␈α
identity␈α
of␈α∞an␈α
organism,
␈↓ α←␈↓↓only␈αuse␈αrules␈αwritten␈αby␈αDr␈αJones␈αor␈αDr␈αSmith␈↓,␈αor␈α␈↓↓...only␈αthose␈αwritten␈αby␈αthe
␈↓ α←␈↓↓expert␈α∂currently␈α∂running␈α∂the␈α∂program␈↓.␈α∂ (There␈α∂might␈α∂even␈α∂be␈α∂a␈α∞preference

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[17]␈αIf␈α
the␈αexpert␈α
never␈αanswers␈αa␈α
question␈αwith␈α
``unknown,''␈αthe␈α
problem␈αis
␈↓ α←␈↓clearly␈αan␈αinsufficiency␈αin␈αthe␈αknowledge␈αbase.␈α All␈αother␈αcases,␈αhowever,␈αare
␈↓ α←␈↓indeterminant and can be a result of both problems.
␈↓ α←␈↓␈↓5-6␈↓ π∞PERSONALIZED WORLD VIEWS    113␈↓

␈↓"β␈↓ α←␈↓ordering␈αon␈αthe␈αrules,␈αbased␈αon␈αlength␈αof␈αexperience␈αof␈αthe␈αrule␈αauthor.␈α See
␈↓ α←␈↓chapter␈α∀7␈α∃for␈α∀the␈α∀details␈α∃of␈α∀meta-rule␈α∀operation.)␈α∃Using␈α∀both␈α∃of␈α∀these
␈↓ α←␈↓capabilities,␈α∂it␈α∂is␈α∂possible␈α∂to␈α∂see␈α∂both␈α∂how␈α∂a␈α∂given␈α∂expert␈α∂reasons␈α⊂about␈α∂a
␈↓ α←␈↓subject and how he might handle a single, given aspect of a case.
␈↓ α←␈↓␈↓ β?While␈αthis␈αprovides␈αa␈αhandle␈αon␈αmanipulating␈αdifferent␈αworld␈αviews,
␈↓ α←␈↓it␈α⊂clearly␈α⊂does␈α⊃not␈α⊂confront␈α⊂the␈α⊂deeper␈α⊃problem␈α⊂of␈α⊂resolving␈α⊃any␈α⊂conflicts
␈↓ α←␈↓between them.  More work on this subject is required.

␈↓ α←␈↓␈↓α5-7    MORE ON MODELS, CONCEPT FORMATION, AND MODEL-
␈↓ α←␈↓α␈↓ β3BASED UNDERSTANDING␈↓
␈↓ α←␈↓␈↓ β?Now␈α
that␈α
the␈α
organization,␈α
structure,␈α
and␈α
function␈α
of␈α
rule␈α
models␈α
in
␈↓ α←␈↓the␈α⊃system␈α⊂is␈α⊃clear,␈α⊃let's␈α⊂take␈α⊃another␈α⊂look␈α⊃at␈α⊃the␈α⊂general␈α⊃idea␈α⊃of␈α⊂models,
␈↓ α←␈↓concept formation, and model-based understanding.

␈↓ α←␈↓␈↓α5-7-1    Model-based understanding␈↓
␈↓ α←␈↓␈↓ β?Many␈αAI␈αprograms␈αhave␈αincorporated␈αexplicit␈αmodels␈αand␈αused␈αthem
␈↓ α←␈↓as␈α
a␈α
guide␈α∞to␈α
understanding.␈α
 The␈α∞work␈α
of␈α
Falk␈α∞and␈α
Roberts␈α
in␈α∞vision␈α
has
␈↓ α←␈↓been␈α⊂mentioned;␈α⊃the␈α⊂idea␈α⊃has␈α⊂been␈α⊂extended␈α⊃to␈α⊂a␈α⊃range␈α⊂of␈α⊃other␈α⊂sensory
␈↓ α←␈↓modalities␈α∀as␈α∃well.␈α∀ Viewed␈α∀in␈α∃the␈α∀broadest␈α∀terms,␈α∃the␈α∀process␈α∃may␈α∀be
␈↓ α←␈↓considered one of signal interpretation, as suggested by the figure below.

␈↓"␈↓ α←␈↓∧                        ⊂αααααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧ ←ααααinterpretationααα ~  signal processor  ~  ←ααααsignalααα
␈↓"␈↓ α←␈↓∧                        %αααααααααααααααααααα$
␈↓"␈↓ α←␈↓∧                                 ↑
␈↓"␈↓ α←␈↓∧     ⊂ααααααααα⊃                 ~
␈↓"␈↓ α←␈↓∧     ~  model  ~ αααααααααααααααα$
␈↓"␈↓ α←␈↓∧     %ααααααααα$

␈↓"␈↓ α←␈↓α␈↓ βTFig. 5-8.    Simple view of model-based understanding.    

␈↓ α←␈↓Some␈α
signal-processing␈α
operation␈αis␈α
performed␈α
on␈αthe␈α
incoming␈α
signal,␈αwith
␈↓ α←␈↓guidance␈αand/or␈αadvice␈αfrom␈αa␈αmodel.␈α The␈αresult␈αis␈αan␈αinterpretation␈αof␈αthe
␈↓ α←␈↓signal␈α∃(or␈α∃some␈α∃action␈α∀based␈α∃on␈α∃it),␈α∃which␈α∀gives␈α∃evidence␈α∃that␈α∃it␈α∀was
␈↓ α←␈↓understood.␈α The␈αtable␈α
below␈αlists␈αsix␈α
different␈αsystems␈αthat␈α
can␈αbe␈αviewed␈α
in
␈↓ α←␈↓these terms (including ours) and compares them along several dimensions.
␈↓ α←␈↓␈↓114    KNOWLEDGE ACQUISITION I␈↓ 
#5-7␈↓


␈↓"β␈↓ α←␈↓α␈↓ ∧aTable 5-1.    Model-based systems.    

␈↓"␈↓ α←␈↓∧αααααααααααααπαααααααααααπααααααααααααααπαααααααπαααααααααααααα
␈↓"␈↓ α←␈↓∧Reference    ~Signal     ~ Model        ~Model  ~ Evidence of
␈↓"␈↓ α←␈↓∧             ~           ~              ~type   ~ understanding
␈↓"␈↓ α←␈↓∧αααααααααααααβαααααααααααβααααααααααααααβαααααααβαααααααααααααα
␈↓"␈↓ α←␈↓∧[Falk70]     ~Light      ~Polyhedra     ~Object ~Synthesized
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~reconstruction
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~of picture
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~
␈↓"␈↓ α←␈↓∧[Winograd72] ~English    ~Blocks world  ~World  ~Block
␈↓"␈↓ α←␈↓∧             ~text       ~              ~       ~manipulation,
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~coherent
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~dialogue
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~
␈↓"␈↓ α←␈↓∧[Waltz72]    ~Line       ~Vertex        ~World  ~Analysis of
␈↓"␈↓ α←␈↓∧             ~drawings   ~classification~       ~picture
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~
␈↓"␈↓ α←␈↓∧[Reddy73]    ~Sound      ~Chess game    ~World  ~Typed version
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~of spoken
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~sentence
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~
␈↓"␈↓ α←␈↓∧[Goldstein74]~Annotated  ~Verbal        ~World  ~Debugged
␈↓"␈↓ α←␈↓∧             ~program    ~description of~       ~program
␈↓"␈↓ α←␈↓∧             ~           ~picture       ~       ~
␈↓"␈↓ α←␈↓∧             ~           ~              ~       ~
␈↓"␈↓ α←␈↓∧ Davis       ~English    ~Rule models,  ~Object ~Retranslated
␈↓"␈↓ α←␈↓∧             ~           ~medical       ~world  ~rule
␈↓"␈↓ α←␈↓∧             ~           ~knowledge base~       ~
␈↓"␈↓ α←␈↓∧ααααααααααααα∀ααααααααααα∀αααααααααααααα∀ααααααα∀αααααααααααααα


␈↓ α←␈↓␈↓ β?The examples illustrate the two sorts of models in use.
␈↓ α←␈↓␈↓ β?``World␈αmodel''␈αis␈αused␈αin␈αthe␈αusual␈αsense␈αand␈αrefers␈αto␈αa␈αcollection␈αof
␈↓ α←␈↓information␈αthat␈α
characterizes␈αthe␈αdomain␈α
of␈αinterest.␈α The␈α
system␈αdescribed
␈↓ α←␈↓in␈α[Reddy73],␈αfor␈αexample,␈α
had␈αa␈αmodel␈αof␈α
the␈αworld␈αof␈αchess␈α
that␈αincluded
␈↓ α←␈↓information␈αabout␈αplausible␈αmoves␈αand␈αappropriate␈αvocabulary.␈α
 The␈αvision
␈↓ α←␈↓system␈α∃in␈α∀[Waltz72]␈α∃contained␈α∀a␈α∃complete␈α∀classification␈α∃of␈α∀the␈α∃types␈α∀of
␈↓ α←␈↓vertices␈αthat␈αcould␈αbe␈αa␈αpart␈αof␈αa␈αconvex␈αplanar-faced␈αpolyhedron␈αand␈αused
␈↓ α←␈↓this as its model of the blocks world.
␈↓ α←␈↓␈↓ β?An␈α⊃``object␈α⊂model,''␈α⊃on␈α⊂the␈α⊃other␈α⊂hand,␈α⊃is␈α⊂used␈α⊃to␈α⊃characterize␈α⊂the
␈↓ α←␈↓expected␈α∂content␈α∞of␈α∂the␈α∞signal␈α∂to␈α∞be␈α∂processed.␈α∞ Falk's␈α∂system,␈α∂for␈α∞example,
␈↓ α←␈↓had␈α⊗models␈α∃of␈α⊗the␈α∃individual␈α⊗polyhedra␈α∃from␈α⊗which␈α∃scenes␈α⊗would␈α∃be
␈↓ α←␈↓constructed.␈α∂ Understanding␈α∂a␈α∞scene␈α∂was␈α∂then␈α∞viewed␈α∂in␈α∂terms␈α∂of␈α∞choosing
␈↓ α←␈↓specific models and assigning locations and orientations to each.
␈↓ α←␈↓␈↓ β?The␈α∂two␈α⊂sorts␈α∂of␈α∂models␈α⊂tend␈α∂to␈α∂be␈α⊂used␈α∂somewhat␈α⊂differently.␈α∂ In
␈↓ α←␈↓very␈α≥general␈α≥terms,␈α≥we␈α≥can␈α≤say␈α≥that␈α≥since␈α≥object␈α≥models␈α≤provide
␈↓ α←␈↓characterizations␈α
of␈α
signal␈α
content,␈α∞``understanding''␈α
can␈α
be␈α
approached␈α∞as␈α
a
␈↓"β␈↓ α←␈↓␈↓5-7␈↓ ∧kMODELS, CONCEPT FORMATION, AND UNDERSTANDING    115␈↓

␈↓"β␈↓ α←␈↓comparison␈α∩between␈α⊃the␈α∩model␈α⊃and␈α∩the␈α⊃signal.␈α∩ World␈α⊃models␈α∩provide␈α⊃a
␈↓ α←␈↓source␈α∃of␈α∃knowledge␈α∀that␈α∃is␈α∃typically␈α∃used␈α∀to␈α∃check␈α∃the␈α∃plausibility␈α∀of
␈↓ α←␈↓potential␈α
interpretations␈α
of␈αthe␈α
signal.␈α
Object␈αmodels␈α
are␈α
thus␈α
typically␈αused
␈↓ α←␈↓in␈α→some␈α→manner␈α→of␈α→matching␈α→process,␈α→while␈α→world␈α→models␈α~are␈α→used
␈↓ α←␈↓inferentially.
␈↓ α←␈↓␈↓ β?The␈αtwo␈αmodels␈αare␈αreally␈αat␈αeither␈αextreme␈αof␈αthe␈αcontinuum␈αshown
␈↓ α←␈↓below,␈αwhich␈αemphasizes␈αthe␈α
distinction␈αby␈αconsidering␈αtwo␈α
vision␈αprograms
␈↓ α←␈↓that␈α↔performed␈α↔similar␈α↔tasks.␈↓
18␈↓␈α↔Where␈α↔Falk's␈α↔program␈α↔had␈α↔models␈α⊗of
␈↓ α←␈↓individual␈α⊗polyhedra␈α∃(models␈α⊗characterizing␈α∃the␈α⊗signal␈α⊗content),␈α∃Waltz's
␈↓ α←␈↓program␈αused␈αknowledge␈αabout␈αthe␈α``world''␈αof␈αpolyhedra␈αin␈αgeneral,␈αwithout
␈↓ α←␈↓reference to any specific instance.

␈↓"β␈↓ α←␈↓	             MODEL           PROCESS        EXAMPLE

␈↓"β␈↓ α←␈↓	          World model       Inferential     [Waltz72]

␈↓"β␈↓ α←␈↓	          Object model      Matching        [Falk70]


␈↓"β␈↓ α←␈↓α␈↓ β<Fig. 5-9.    Object models and world models:  Comparison.    

␈↓ α←␈↓␈↓ β?The␈α
rule␈α
models␈α∞used␈α
in␈α
␈↓¬TEIRESIAS␈↓␈α∞fall␈α
somewhere␈α
in␈α∞between.␈α
 They
␈↓ α←␈↓are␈α
a␈α
type␈α
of␈α∞object␈α
model,␈α
since␈α
they␈α∞represent␈α
the␈α
content␈α
expected␈α∞in␈α
the
␈↓ α←␈↓signal.␈α⊂ But␈α⊂they␈α∂are␈α⊂not␈α⊂themselves␈α⊂rules␈α∂and␈α⊂cannot␈α⊂therefore␈α⊂simply␈α∂be
␈↓ α←␈↓matched␈α⊂against␈α⊂the␈α∂signal.␈α⊂ The␈α⊂information␈α∂they␈α⊂carry,␈α⊂however,␈α⊂can␈α∂be
␈↓ α←␈↓used␈α⊂to␈α⊂help␈α⊂direct␈α⊂the␈α∂interpretation␈α⊂process.␈α⊂ In␈α⊂addition,␈α⊂the␈α∂knowledge
␈↓ α←␈↓base␈α∞provides␈α∞the␈α∞system␈α∞with␈α∞a␈α
world␈α∞model.␈α∞ As␈α∞will␈α∞become␈α∞clear␈α∞in␈α
the
␈↓ α←␈↓next␈αchapter,␈αthe␈αknowledge␈αbase␈αcontains␈αmany␈αdata␈αstructures␈αthat␈αindicate
␈↓ α←␈↓such␈αthings␈αas␈αwhich␈αvalues␈αbelong␈αwith␈αwhich␈αattributes.␈α The␈αsystem␈αrelies
␈↓ α←␈↓on␈α⊃this␈α⊃information␈α⊃to␈α⊃maintain␈α⊃the␈α⊃internal␈α⊃consistency␈α⊃of␈α⊃the␈α⊃clauses␈α⊂it
␈↓ α←␈↓assembles␈αand␈αwill,␈αas␈α
a␈αresult,␈αnever␈αproduce␈αa␈α
clause␈αof␈αthe␈αform␈α``␈↓	The␈α
site
␈↓ α←␈↓	of the culture is e.coli.␈↓''

␈↓ α←␈↓␈↓α5-7-2    Concept formation␈↓
␈↓ α←␈↓␈↓ β?There␈αhas␈αalso␈α
been␈αmuch␈αwork␈αdone␈α
in␈αAI␈αon␈αconcept␈α
formation;␈αa
␈↓ α←␈↓recent␈αexample␈αnoted␈αearlier␈αis␈α[Winston70].␈α A␈αmuch␈αsimplified␈αview␈αof␈αthe
␈↓ α←␈↓information␈αflow␈αin␈αthe␈αtask␈αis␈αshown␈αbelow,␈αwhere␈αa␈αset␈αof␈αexamples␈αis␈αused
␈↓ α←␈↓as the basis for inferring a concept.





␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[18]␈αRemember␈α
that␈αselected␈α
aspects␈αof␈α
each␈αprogram␈α
have␈αbeen␈α
singled␈αout
␈↓ α←␈↓here␈αfor␈αthe␈αsake␈αof␈αdiscussion.␈α
 All␈αof␈αthe␈αprograms␈αhere␈αand␈αin␈α
Table␈α5-1
␈↓ α←␈↓were more complex than the simple characterizations given.
␈↓ α←␈↓␈↓116    KNOWLEDGE ACQUISITION I␈↓ 
#5-7␈↓


␈↓"␈↓ α←␈↓∧                  ⊂αααααααααα⊃
␈↓"␈↓ α←␈↓∧                  ~ examples ~
␈↓"␈↓ α←␈↓∧                  %αααααααααα$
␈↓"␈↓ α←␈↓∧                      ~                   ⊂αααααααααα⊃
␈↓"␈↓ α←␈↓∧                      %ααααααααααα→       ~  concept ~
␈↓"␈↓ α←␈↓∧                        concept           %αααααααααα$
␈↓"␈↓ α←␈↓∧                        formation

␈↓"␈↓ α←␈↓α␈↓ βoFig. 5-10.    Simplified view of concept formation.    


␈↓ α←␈↓␈↓α5-7-3    A synthesis␈↓
␈↓ α←␈↓␈↓ β?Now␈αconsider␈αthe␈α
process␈αof␈αmodel-directed␈α
understanding␈αcombined
␈↓ α←␈↓with␈αconcept␈αformation,␈αproducing␈αthe␈αfigure␈αbelow.␈α In␈αterms␈αof␈αour␈αsystem,
␈↓ α←␈↓the␈α⊃``signal''␈α⊃is␈α⊃the␈α∩new␈α⊃rule␈α⊃text,␈α⊃``signal␈α∩processing''␈α⊃is␈α⊃done␈α⊃by␈α∩the␈α⊃rule
␈↓ α←␈↓acquisition␈αroutines,␈αthe␈αinterpretation␈αis␈αthe␈αnew␈αrule␈αin␈α␈↓¬LISP␈↓␈αterms,␈αand␈αthe
␈↓ α←␈↓``model''␈α∂is␈α∞provided␈α∂by␈α∞the␈α∂set␈α∂of␈α∞rule␈α∂models.␈α∞ For␈α∂concept␈α∂formation,␈α∞the
␈↓ α←␈↓rule␈αbase␈αsupplies␈αthe␈αexamples␈αfrom␈αwhich␈αthe␈αconcepts--the␈α
rule␈αmodels--
␈↓ α←␈↓are constructed.


␈↓"␈↓ α←␈↓∧⊂ααααααααααα⊃    [knowledge     ⊂ααααααααααα⊃
␈↓"␈↓ α←␈↓∧~ KNOWLEDGE ~ ← α α α α α α α α ~   RULE    ~ ← α α α α EXPERT
␈↓"␈↓ α←␈↓∧~   BASE    ~    acquisition]   ~ACQUISITION~  [dialog]
␈↓"␈↓ α←␈↓∧%ααααααααααα$                   %ααααααααααα$
␈↓"␈↓ α←␈↓∧    ~                               ↑

␈↓"␈↓ α←␈↓∧    ~                               ~

␈↓"␈↓ α←␈↓∧    ~          ⊂αααααααααα⊃         ~
␈↓"␈↓ α←␈↓∧               ~   RULE   ~           [model-directed
␈↓"␈↓ α←␈↓∧    % α α α α →~  MODELS  ~ α α α α $  understanding]
␈↓"␈↓ α←␈↓∧[concept       %αααααααααα$
␈↓"␈↓ α←␈↓∧ formation]

␈↓ α←␈↓α␈↓ β≠Fig.␈α⊗5-11.    Synthesis␈α↔of␈α⊗concept␈α⊗formation␈α↔and␈α⊗model-based
␈↓ α←␈↓α␈↓ β≠understanding.    

␈↓ α←␈↓␈↓ β?The␈α∀result␈α∃is␈α∀an␈α∀interesting␈α∃form␈α∀of␈α∀closed-loop␈α∃behavior.␈α∀ The
␈↓ α←␈↓existing␈α
rule␈α
models␈α
are␈αused␈α
to␈α
guide␈α
the␈αprocess␈α
of␈α
acquisition,␈α
the␈αnew␈α
rule
␈↓ α←␈↓is␈αadded␈αto␈α
the␈αknowledge␈αbase␈α
and␈αthe␈αrelevant␈α
rule␈αmodels␈αare␈α
recomputed.
␈↓ α←␈↓The␈αsystem␈αis␈αthus␈αconstructing␈α
its␈αmodels␈α(its␈αpicture␈αof␈αits␈α
own␈αknowledge)
␈↓ α←␈↓on␈αthe␈αbasis␈αof␈αexperience,␈αkeeping␈αthose␈αmodels␈αup-to-date␈αwith␈αthe␈αcurrent
␈↓ α←␈↓knowledge base and using them to help acquire new knowledge.
␈↓ α←␈↓␈↓ β?This␈α≡loop␈α≡has␈α≡a␈α≡number␈α≡of␈α≡interesting␈α∨implications.␈α≡ First,
␈↓ α←␈↓performance␈α∞on␈α∞the␈α∞acquisition␈α∞of␈α∞the␈α∞next␈α∞rule␈α∞may␈α∞be␈α∞better␈α∂because␈α∞the
␈↓ α←␈↓system's␈α∞``picture''␈α∞of␈α∞its␈α∞knowledge␈α
base␈α∞has␈α∞improved--the␈α∞rule␈α∞models␈α
are
␈↓"β␈↓ α←␈↓␈↓5-7␈↓ ∧kMODELS, CONCEPT FORMATION, AND UNDERSTANDING    117␈↓

␈↓"β␈↓ α←␈↓now␈α∂computed␈α∂from␈α∂a␈α∂larger␈α∂set␈α∂of␈α∂instances␈α∂and␈α∂their␈α∂generalizations␈α∂are
␈↓ α←␈↓more likely to be valid.
␈↓ α←␈↓␈↓ β?Second,␈α∂since␈α⊂the␈α∂relevant␈α⊂rule␈α∂models␈α∂are␈α⊂recomputed␈α∂each␈α⊂time␈α∂a
␈↓ α←␈↓change␈α∪is␈α∪made␈α∪to␈α∪the␈α∪knowledge␈α∩base,␈α∪the␈α∪picture␈α∪they␈α∪supply␈α∪is␈α∩kept
␈↓ α←␈↓constantly␈αup-to-date;␈αthus,␈αthey␈αare␈αat␈αall␈αtimes␈αan␈αaccurate␈αreflection␈αof␈αthe
␈↓ α←␈↓shifting patterns in the knowledge base.
␈↓ α←␈↓␈↓ β?Finally,␈αand␈αperhaps␈αmost␈αinteresting,␈αthe␈αmodels␈αare␈αnot␈αhand-tooled
␈↓ α←␈↓by␈αthe␈α
system␈αarchitect␈αor␈α
specified␈αby␈αthe␈α
expert.␈α They␈αare␈α
instead␈αformed
␈↓ α←␈↓by␈α
the␈α
system␈α
itself␈α
and␈α
formed␈α
as␈α
a␈α
result␈α
of␈α
its␈α
experience␈α
in␈αacquiring␈α
rules
␈↓ α←␈↓from␈αthe␈αexpert.␈α Thus,␈αdespite␈αits␈αreliance␈αon␈αa␈αset␈αof␈αmodels␈αas␈αa␈α
basis␈αfor
␈↓ α←␈↓understanding,␈α⊂␈↓¬TEIRESIAS␈↓'s␈α⊂abilities␈α⊃are␈α⊂not␈α⊂restricted␈α⊃by␈α⊂the␈α⊂existing␈α⊃set␈α⊂of
␈↓ α←␈↓models.␈α
 As␈α
its␈α
store␈α
of␈α
knowledge␈α
grows,␈α
old␈α
models␈α
become␈α
more␈αaccurate,
␈↓ α←␈↓new␈α∀models␈α∀are␈α∪formed,␈α∀and␈α∀the␈α∀system's␈α∪stock␈α∀of␈α∀knowledge␈α∀about␈α∪its
␈↓ α←␈↓knowledge␈αcontinues␈αto␈αexpand.␈α This␈αappears␈αto␈αbe␈αa␈αnovel␈αcapability␈αfor␈αa
␈↓ α←␈↓model-based system.
␈↓ α←␈↓␈↓118    KNOWLEDGE ACQUISITION I␈↓ 
#5-8␈↓

␈↓"β␈↓ α←␈↓␈↓α5-8    UNSOLVED PROBLEMS AND FUTURE WORK␈↓
␈↓ α←␈↓␈↓ β?There␈α⊗are␈α∃a␈α⊗number␈α⊗of␈α∃problems␈α⊗left␈α∃unsolved␈α⊗by␈α⊗the␈α∃current
␈↓ α←␈↓implementation␈αthat␈αsuggest␈α
several␈αdirections␈αfor␈α
future␈αwork.␈α They␈α
are␈αof
␈↓ α←␈↓two␈α≤forms:␈α≤minor␈α≤problems␈α≤whose␈α≤solution␈α≤involves␈α≤extensions␈α≠and
␈↓ α←␈↓refinements␈α~to␈α~existing␈α~methods,␈α~and␈α~major␈α~problems␈α~requiring␈α→new
␈↓ α←␈↓solutions.

␈↓ α←␈↓␈↓α5-8-1    Minor problems␈↓
␈↓ α←␈↓␈↓ β?As␈α∞with␈α∞all␈α∞first-generation␈α∞systems,␈α∞␈↓¬TEIRESIAS␈↓␈α∞has␈α∞many␈α∂rough␈α∞spots
␈↓ α←␈↓and␈α↔inconveniences.␈α⊗ These␈α↔will␈α↔have␈α⊗to␈α↔be␈α⊗smoothed␈α↔out␈α↔before␈α⊗the
␈↓ α←␈↓acquisition␈α↔routines␈α↔can␈α↔become␈α↔a␈α↔powerful␈α↔user-oriented␈α↔system.␈α⊗ For
␈↓ α←␈↓instance,␈αthe␈α
system's␈αguidance␈α
of␈αthe␈αdebugging␈α
process␈αcould␈α
be␈αimproved.
␈↓ α←␈↓It␈α∞currently␈α∂derives␈α∞much␈α∂power␈α∞from␈α∞its␈α∂methodical␈α∞approach,␈α∂forcing␈α∞the
␈↓ α←␈↓expert's␈α∞criticism␈α∂to␈α∞be␈α∂sharply␈α∞focused.␈α∂ But␈α∞it␈α∂is␈α∞also␈α∂somewhat␈α∞inflexible
␈↓ α←␈↓and␈αunforgiving: ␈αIn␈αmost␈αcases␈αthere␈αis␈αno␈αway␈αto␈αchange␈αa␈αresponse␈αonce␈αa
␈↓ α←␈↓question is answered, for instance.
␈↓ α←␈↓␈↓ β?There␈α∀are␈α∀other␈α∀tasks␈α∀suggested␈α∀by␈α∀the␈α∀larger␈α∀context␈α∃in␈α∀which
␈↓ α←␈↓acquisition␈α
occurs.␈α
 It␈αwould␈α
be␈α
useful,␈α
for␈αexample,␈α
to␈α
provide␈α
several␈αsorts
␈↓ α←␈↓of␈α
feedback␈α∞on␈α
the␈α∞consultation␈α
system's␈α
performance.␈α∞ A␈α
recent␈α∞addition␈α
to
␈↓ α←␈↓␈↓¬MYCIN␈↓␈αkeeps␈αextensive␈αstatistics␈αon␈αthe␈αuse␈αof␈αeach␈αrule␈αin␈αthe␈αknowledge␈αbase.
␈↓ α←␈↓These␈α∃should␈α∃be␈α∃routinely␈α∃scanned␈α∀by␈α∃the␈α∃acquisition␈α∃system␈α∃to␈α∀detect
␈↓ α←␈↓potential␈αbugs␈α(e.g.,␈αa␈αrule␈αthat␈αis␈αnever␈αinvoked␈αsuccessfully␈αis␈αlikely␈αto␈αhave
␈↓ α←␈↓an␈α∪error␈α∀in␈α∪its␈α∀premise).␈α∪ A␈α∀more␈α∪sophisticated␈α∀solution␈α∪might␈α∀even␈α∪be
␈↓ α←␈↓capable␈α
of␈α
suggesting␈α
plausible␈α∞corrections,␈α
based␈α
on␈α
an␈α
examination␈α∞of␈α
the
␈↓ α←␈↓situations␈α
under␈α
which␈α
failure␈α
occurred.␈α
 It␈α
should␈α
also␈α
be␈α
possible␈αto␈α
provide
␈↓ α←␈↓some␈α↔feedback␈α⊗after␈α↔the␈α↔complete␈α⊗information␈α↔on␈α⊗a␈α↔case␈α↔is␈α⊗available.
␈↓ α←␈↓Original␈α⊃diagnoses,␈α⊃for␈α⊃instance,␈α⊃could␈α⊃be␈α⊃evaluated␈α⊃in␈α⊃light␈α⊃of␈α⊃the␈α⊃final
␈↓ α←␈↓results␈α≠from␈α≠the␈α≠laboratory,␈α≠possibly␈α≠suggesting␈α≠modifications␈α≠to␈α≠the
␈↓ α←␈↓knowledge base.
␈↓ α←␈↓␈↓ β?The␈αrule␈αeditor␈αcould␈αalso␈αbe␈αimproved␈αin␈αseveral␈αrespects.␈α It␈αshould
␈↓ α←␈↓be␈αpossible,␈αfor␈αinstance,␈αto␈αmake␈αa␈α
number␈αof␈αroutine␈αchanges␈αto␈αa␈αrule␈α
with
␈↓ α←␈↓less␈α∩machinery␈α∩than␈α∩is␈α∩currently␈α∩used.␈α∩ The␈α∩rule␈α∩editor␈α∩can␈α∩currently␈α⊃be
␈↓ α←␈↓invoked␈αseparately␈αto␈αaccomplish␈αsome␈αof␈αthese,␈αbut␈αit␈αwould␈αrequire␈αa␈α
larger
␈↓ α←␈↓command␈α∞set␈α∞to␈α∞be␈α∞adequately␈α∞powerful.␈α∞ A␈α∞more␈α∞substantive␈α∞improvement
␈↓ α←␈↓concerns␈α
the␈α
problem␈α
noted␈α
earlier,␈α
of␈α
knowing␈α
exactly␈α
what␈α
to␈α∞change␈α
and
␈↓ α←␈↓what␈α
to␈α∞delete␈α
in␈α∞a␈α
rule␈α
that␈α∞has␈α
been␈α∞misunderstood.␈α
 While␈α∞the␈α
primitive
␈↓ α←␈↓approach␈αto␈αnatural␈αlanguage␈αis␈α
the␈αfundamental␈αsource␈αof␈αthe␈α
problem␈α(see
␈↓ α←␈↓below),␈αthere␈αare␈αseveral␈αthings␈αthat␈αcould␈αbe␈αdone␈αto␈αease␈αthe␈αdifficulties.␈α It
␈↓ α←␈↓would␈αbe␈αa␈αgreat␈αhelp␈αsimply␈αto␈αmake␈αclear␈αwhich␈αinterpretations␈αcame␈αfrom
␈↓ α←␈↓which␈α⊃lines␈α⊃of␈α⊃the␈α⊃original␈α⊃input␈α∩text.␈α⊃ This␈α⊃might␈α⊃be␈α⊃done␈α⊃as␈α∩easily␈α⊃as
␈↓ α←␈↓grouping␈α∩the␈α∩appropriate␈α∩lines␈α∪together␈α∩as␈α∩they␈α∩are␈α∩printed,␈α∪making␈α∩the
␈↓ α←␈↓nature of the system's misunderstandings more obvious.
␈↓ α←␈↓␈↓ β?Since␈α∀the␈α∀process␈α∀of␈α∀tracking␈α∀down␈α∀the␈α∀original␈α∀problem␈α∀in␈α∀the
␈↓ α←␈↓knowledge␈α∪base␈α∀is␈α∪easily␈α∀viewed␈α∪as␈α∀diagnosis␈α∪and␈α∀therapy,␈α∪there␈α∀is␈α∪the
␈↓ α←␈↓interesting␈α⊂possibility␈α⊂of␈α⊃expressing␈α⊂it␈α⊂too␈α⊂in␈α⊃rules.␈α⊂ Such␈α⊂a␈α⊂body␈α⊃of␈α⊂rules
␈↓ α←␈↓␈↓5-8␈↓ ¬|UNSOLVED PROBLEMS AND FUTURE WORK    119␈↓

␈↓"β␈↓ α←␈↓would␈α∪allow␈α∪running␈α∪a␈α∀``mini-consultation''␈α∪to␈α∪uncover␈α∪the␈α∀problem␈α∪and
␈↓ α←␈↓initiate␈α⊗corrective␈α⊗action.␈α⊗ It␈α∃would␈α⊗have␈α⊗the␈α⊗substantive␈α⊗advantage␈α∃of
␈↓ α←␈↓allowing␈α
all␈α
the␈α
explanation␈α
(and␈α
conceivably,␈α
acquisition)␈α
routines␈α
to␈α
be␈α
used
␈↓ α←␈↓during␈αthe␈αdebugging␈αprocess.␈α Rules␈αto␈αdo␈αthis␈αhave␈αbeen␈αdrafted␈αbut␈αhave
␈↓ α←␈↓not yet been implemented.

␈↓ α←␈↓␈↓α5-8-2    Major problems␈↓

␈↓ α←␈↓␈↓αBetter techniques for rule model generation␈↓    
␈↓ α←␈↓␈↓ β?The␈α∞shortcomings␈α∞of␈α∞the␈α∞present␈α∞approach␈α∞to␈α∞rule␈α∞model␈α
generation
␈↓ α←␈↓were␈αoutlined␈αearlier.␈α The␈αprimary␈αproblem␈αis␈αthe␈αuse␈αof␈αa␈αpurely␈αstatistical
␈↓ α←␈↓approach␈α∂to␈α∞concept␈α∂formation.␈α∂ This␈α∞approach␈α∂was␈α∞motivated,␈α∂in␈α∂part,␈α∞by
␈↓ α←␈↓the␈αdesire␈αto␈αmake␈αthe␈αmodels␈αtransparent␈αto␈αthe␈αexpert.␈α More␈αsophisticated
␈↓ α←␈↓sorts␈α
of␈α
concept␈α
formation␈α
would␈α
be␈α
possible␈α
if␈α
the␈α
model␈α
construction␈α
process
␈↓ α←␈↓were␈α∂made␈α∞interactive.␈α∂ With␈α∞this␈α∂approach,␈α∞each␈α∂time␈α∞a␈α∂change␈α∂had␈α∞been
␈↓ α←␈↓made␈αto␈α
the␈αknowledge␈α
base,␈α␈↓¬TEIRESIAS␈↓␈αwould␈α
indicate␈αto␈α
the␈αexpert␈α
any␈αnew
␈↓ α←␈↓patterns␈α∞that␈α∞had␈α
become␈α∞evident␈α∞and␈α
would␈α∞ask␈α∞him␈α
for␈α∞an␈α∞evaluation␈α
of
␈↓ α←␈↓each.␈α∩ With␈α∩this␈α∪sort␈α∩of␈α∩advice,␈α∩it␈α∪would␈α∩become␈α∩possible␈α∪to␈α∩distinguish
␈↓ α←␈↓accidental␈αcorrelations␈αfrom␈αvalid␈αinterrelations.␈α It␈αwould␈αalso␈αbe␈αpossible␈αto
␈↓ α←␈↓construct␈α∪models␈α∪with␈α∪a␈α∪much␈α∪finer␈α∪degree␈α∪of␈α∪control.␈α∪ The␈α∀utility␈α∪and
␈↓ α←␈↓sophistication of ␈↓¬TEIRESIAS␈↓'s second guessing would increase proportionally.

␈↓ α←␈↓␈↓αNatural language␈↓    
␈↓ α←␈↓␈↓ β?Of␈α↔the␈α↔major␈α⊗problems,␈α↔the␈α↔weakness␈α⊗of␈α↔the␈α↔natural␈α⊗language
␈↓ α←␈↓understanding␈α
techniques␈α
presents␈α∞the␈α
largest␈α
barrier␈α
to␈α∞better␈α
performance.
␈↓ α←␈↓Even␈α∂without␈α∞introducing␈α∂more␈α∞sophisticated␈α∂grammar-oriented␈α∞techniques,
␈↓ α←␈↓however,␈αthere␈α
are␈αseveral␈α
steps␈αthat␈α
might␈αbe␈α
taken␈αto␈α
strengthen␈αthe␈α
system.
␈↓ α←␈↓Processing␈α⊂each␈α⊂line␈α⊃of␈α⊂the␈α⊂text␈α⊂independently␈α⊃is␈α⊂one␈α⊂source␈α⊃of␈α⊂weakness.
␈↓ α←␈↓The␈α∞rule␈α∞models␈α∞are␈α∞currently␈α∂used␈α∞to␈α∞score␈α∞the␈α∞interpretation␈α∞of␈α∂each␈α∞line
␈↓ α←␈↓independently,␈α
but␈α
the␈α
ntuples␈α
might␈αeasily␈α
be␈α
used␈α
to␈α
score␈α
an␈αentire␈α
premise
␈↓ α←␈↓at once.
␈↓ α←␈↓␈↓ β?Also,␈αas␈αshown,␈αthere␈αis␈αsubstantive␈αadvantage␈αin␈αbeing␈αcareful␈αabout
␈↓ α←␈↓the␈α∞consistency␈α∞of␈α∞the␈α∞parses␈α∞generated␈α∂for␈α∞a␈α∞single␈α∞line␈α∞of␈α∞text.␈α∂ A␈α∞similar
␈↓ α←␈↓technique␈αshould␈αbe␈αdeveloped␈αto␈αconsider␈αconsistency␈αbetween␈αlines.␈α This␈α
is
␈↓ α←␈↓a␈αmuch␈αharder␈αproblem,␈αhowever,␈αsince␈αit␈αrequires␈αa␈αconsiderably␈αlarge␈αstore
␈↓ α←␈↓of␈α∞world␈α
knowledge.␈α∞ For␈α∞instance,␈α
it␈α∞makes␈α∞sense␈α
to␈α∞have␈α∞a␈α
rule␈α∞say␈α∞␈↓↓if␈α
the
␈↓ α←␈↓↓identity␈α
of␈α
an␈α
organism␈α
is␈α
likely␈α
to␈α
be␈α
X␈α
and␈α
likely␈α
to␈α
be␈α
Y␈↓␈α
(i.e.,␈α
there's␈α
evidence
␈↓ α←␈↓for␈α⊂both␈α∂X␈α⊂and␈α⊂Y).␈α∂ But␈α⊂it␈α⊂does␈α∂not␈α⊂make␈α⊂sense␈α∂to␈α⊂say␈α⊂␈↓↓if␈α∂the␈α⊂site␈α⊂of␈α∂the
␈↓ α←␈↓↓culture␈α∂is␈α∂likely␈α∂to␈α∂be␈α∂X␈α∂and␈α∂likely␈α∂to␈α∂be␈α∂Y␈↓,␈α∂because␈α∂the␈α∂site␈α∂from␈α∂which␈α∂a
␈↓ α←␈↓culture␈α∞is␈α
obtained␈α∞is␈α∞rarely␈α
in␈α∞doubt.␈α∞This␈α
requires␈α∞more␈α∞knowledge␈α
about
␈↓ α←␈↓things␈αlike␈α
identities␈αand␈α
sites␈αthan␈α
currently␈αexists␈α
in␈αthe␈α
system,␈αor␈α
than␈αis
␈↓ α←␈↓easily added.
␈↓ α←␈↓␈↓ β?Another␈α
manifestation␈α∞of␈α
the␈α
weakness␈α∞of␈α
the␈α∞line-by-line␈α
treatment
␈↓ α←␈↓of␈αnatural␈αlanguage␈αappears␈αin␈αthe␈αway␈αthe␈αEnglish␈αversions␈αof␈αthe␈αrules␈αare
␈↓ α←␈↓printed.␈α⊃ Some␈α⊃rules␈α⊂are␈α⊃quite␈α⊃awkward␈α⊃when␈α⊂translated␈α⊃line␈α⊃by␈α⊃line,␈α⊂yet
␈↓ α←␈↓␈↓120    KNOWLEDGE ACQUISITION I␈↓ 
#5-8␈↓

␈↓"β␈↓ α←␈↓often␈αthey␈αhave␈αvery␈αsimple␈αrestatements␈αin␈αother␈αterms.␈α More␈αsophisticated
␈↓ α←␈↓translation routines should be developed to handle an entire rule at once.
␈↓ α←␈↓␈↓ β?There␈αare␈αobvious␈αweaknesses,␈αtoo,␈αin␈αthe␈αword-by-word␈αapproach␈αto
␈↓ α←␈↓meaning.␈α∃ As␈α∃the␈α∃knowledge␈α∃base␈α∃grows␈α∃larger,␈α∃significant␈α⊗numbers␈α∃of
␈↓ α←␈↓attributes␈αare␈αbeginning␈αto␈α
use␈αsimilar␈αterms.␈αThe␈α
appearance␈αof␈αthat␈αterm␈α
in
␈↓ α←␈↓text␈α∞then␈α
yields␈α∞numerous␈α
connotations␈α∞and␈α
an␈α∞awkwardly␈α
large␈α∞number␈α
of
␈↓ α←␈↓clauses are generated.
␈↓ α←␈↓␈↓ β?The␈α⊂rule␈α⊂models␈α⊂should␈α⊂also␈α∂be␈α⊂integrated␈α⊂deeper␈α⊂into␈α⊂the␈α∂natural
␈↓ α←␈↓language␈α∂interpretation␈α∞process.␈α∂ Rather␈α∞than␈α∂generating␈α∞all␈α∂the␈α∂parses,␈α∞the
␈↓ α←␈↓models␈α⊂should␈α⊂be␈α⊂used␈α⊂to␈α⊂help␈α⊂choose␈α⊂which␈α⊂branch␈α⊂of␈α⊂the␈α⊂parse␈α⊂tree␈α⊂to
␈↓ α←␈↓explore.␈α The␈αtree␈αwould␈α
be␈αgenerated␈αa␈αpath␈α
at␈αa␈αtime,␈αunder␈α
the␈αguidance
␈↓ α←␈↓of␈α⊃the␈α⊃rule␈α⊃model,␈α⊃which␈α⊃might␈α⊃be␈α⊃far␈α⊃more␈α⊃efficient␈α⊃and␈α⊃have␈α⊃a␈α⊂better
␈↓ α←␈↓chance of arriving at the right answer sooner.

␈↓ α←␈↓␈↓αImpact on the knowledge base␈↓    
␈↓ α←␈↓␈↓ β?There␈α
is␈α
a␈α
rather␈α
vast␈α∞problem,␈α
which␈α
we␈α
have␈α
examined␈α∞only␈α
very
␈↓ α←␈↓briefly,␈αconcerning␈αthe␈αimpact␈αof␈αany␈αnew␈αor␈αchanged␈αrule␈αon␈αthe␈αrest␈αof␈αthe
␈↓ α←␈↓knowledge␈α!base.␈↓
19␈↓␈α"There␈α!are␈α!two␈α"general␈α!classes␈α"of␈α!interactions,
␈↓ α←␈↓corresponding␈αto␈αsyntactic␈αand␈α
semantic␈αconflicts.␈α The␈αdifficulty␈α
in␈αsyntactic
␈↓ α←␈↓problems␈α∂arises␈α∂out␈α⊂of␈α∂the␈α∂use␈α∂of␈α⊂certainty␈α∂factors.␈α∂ Except␈α∂in␈α⊂very␈α∂simple
␈↓ α←␈↓cases␈α(e.g.,␈αtwo␈αrules␈αidentical␈αexcept␈αfor␈αtheir␈αCFs),␈αthere␈αis␈αsome␈αquestion␈αof
␈↓ α←␈↓what␈α∂constitutes␈α∂contradiction␈α∂and␈α∂subsumption␈α∂for␈α∂this␈α∂form␈α∂of␈α∂reasoning
␈↓ α←␈↓rule.
␈↓ α←␈↓␈↓ β?The␈α∪lack␈α∪of␈α∩a␈α∪precise␈α∪definition␈α∩of␈α∪inconsistency␈α∪makes␈α∩detecting
␈↓ α←␈↓indirect␈α⊃contradictions␈α⊃(which␈α⊃result␈α⊃from␈α⊃chaining␈α⊃several␈α∩rules␈α⊃together)
␈↓ α←␈↓especially␈αdifficult.␈α As␈αan␈αexample,␈αconsider␈αthe␈αrules␈αbelow.␈αThey␈αare␈αvalid
␈↓ α←␈↓under␈α∪the␈α∪current␈α∀CF␈α∪model␈α∪but␈α∪would␈α∀be␈α∪a␈α∪contradiction␈α∀in␈α∪ordinary
␈↓ α←␈↓binary-valued␈α⊂logic.␈α∂ (A␈α⊂well-specified␈α⊂set␈α∂of␈α⊂organisms␈α∂belong␈α⊂to␈α⊂a␈α∂given
␈↓ α←␈↓category,␈αso␈αjust␈αknowing␈αthe␈αcategory␈αof␈αan␈αorganism␈αgives␈αa␈αhint␈α
about␈αits
␈↓ α←␈↓identity.␈α But␈αthe␈αrule␈αset␈αbelow␈αis␈αplausibly␈αcorrect␈αif␈αall␈αthe␈αmembers␈αof␈αthe
␈↓ α←␈↓category except one [identity A] have the same aerobicity.)

␈↓"β␈↓ α←␈↓	R1:␈↓ ∧∪CATEGORY is X  ␈↓∧===␈↓	 .4 ␈↓∧===@␈↓	 IDENTITY is A
␈↓ α←␈↓␈↓ ∧∪[Knowing␈α↔category␈α↔is␈α↔X␈α↔gives␈α↔some␈α_hint␈α↔of
␈↓ α←␈↓␈↓ ∧∪identity.]

␈↓"β␈↓ α←␈↓	R2:␈↓ ∧∪IDENTITY is A  ␈↓∧===␈↓	 -1 ␈↓∧===@␈↓	 AEROBICITY is B
␈↓ α←␈↓␈↓ ∧∪[Knowing␈α
identity␈α
is␈α
A␈αcan␈α
be␈α
used␈α
definitely␈αto
␈↓ α←␈↓␈↓ ∧∪rule␈α⊂out␈α⊂the␈α⊂aerobicity␈α⊂value␈α⊂which␈α⊂is␈α⊂different
␈↓ α←␈↓␈↓ ∧∪than the aerobicity of A.]




␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[19] See [Shortliffe76] for additional thoughts on this topic.
␈↓ α←␈↓␈↓5-8␈↓ ¬|UNSOLVED PROBLEMS AND FUTURE WORK    121␈↓

␈↓"β␈↓ α←␈↓	R3:␈↓ ∧∪CATEGORY is X  ␈↓∧===␈↓	 .2 ␈↓∧===@␈↓	 AEROBICITY is B
␈↓ α←␈↓␈↓ ∧∪[But␈α∪the␈α∀category␈α∪is␈α∪evidence␈α∀about␈α∪aerobicity
␈↓ α←␈↓␈↓ ∧∪also,␈α∩and␈α∩might␈α∩indicate␈α∩that␈α∩it's␈α∩B,␈α∩if␈α∪all␈α∩the
␈↓ α←␈↓␈↓ ∧∪other members of the category have aerobicity B.]

␈↓ α←␈↓Thus, there are two patterns of the form

␈↓"β␈↓ α←␈↓	␈↓ β7CATEGORY  ␈↓∧==␈↓	 .4 ␈↓∧=@␈↓	  IDENTITY  ␈↓∧==␈↓	 -1.0 ␈↓∧=@␈↓	  AEROBICITY

␈↓ α←␈↓and

␈↓"β␈↓ α←␈↓	␈↓ ∧gCATEGORY  ␈↓∧==␈↓	 .2 ␈↓∧=@␈↓	  AEROBICITY

␈↓ α←␈↓In␈α∀this␈α∀way,␈α∃␈↓	CATEGORY␈↓␈α∀can␈α∀be␈α∀both␈α∃evidence␈α∀in␈α∀favor␈α∀of␈α∃and␈α∀against
␈↓ α←␈↓␈↓	AEROBICITY␈↓,␈α
depending␈α∞on␈α
the␈α∞reasoning␈α
chain␈α∞used.␈↓
20␈↓␈α
Now␈α∞consider␈α
what
␈↓ α←␈↓happens␈αif␈αthe␈αCFs␈αfor␈αR1␈αand␈αR3␈αabove␈αare␈αchanged␈αslowly␈αand␈αallowed␈αto
␈↓ α←␈↓approach␈α
1.0.␈α Since␈α
the␈αCF␈α
model␈αbecomes␈α
Boolean␈αlogic␈α
when␈αall␈α
CFs␈αare
␈↓ α←␈↓either␈α
1.0␈α
or␈α
-1.0,␈αat␈α
some␈α
point␈α
the␈αrule␈α
set␈α
above␈α
will␈α
become␈αinconsistent;
␈↓ α←␈↓but the question is when.
␈↓ α←␈↓␈↓ β?Boolean␈α
logic␈α∞has␈α
the␈α
constraint␈α∞that␈α
P(H|E)␈α
+␈α∞P(not-H|E)␈α
=␈α∞1.0,␈α
an
␈↓ α←␈↓equality␈α
from␈α∞which␈α
the␈α
value␈α∞of␈α
either␈α
probability␈α∞can␈α
be␈α∞computed␈α
when
␈↓ α←␈↓given␈α
the␈α
other.␈α
 For␈α
CFs␈α
there␈α
may␈α
conceivably␈α
be␈α
an␈α
inequality␈αfrom␈α
which
␈↓ α←␈↓to␈α
compute␈α
the␈α
legal␈α
range␈α
of␈α
a␈α
certainty␈α
factor␈α
for␈α
one␈α
rule␈α
in␈α
a␈αparticular
␈↓ α←␈↓set␈α∞of␈α
rules,␈α∞but␈α∞it␈α
has␈α∞not␈α
been␈α∞derived␈α∞as␈α
yet.␈α∞ The␈α
first␈α∞step␈α∞in␈α
detecting
␈↓ α←␈↓inconsistencies is thus to define the concept more precisely.
␈↓ α←␈↓␈↓ β?Even␈α∨with␈α∨a␈α∨rigorous␈α∨definition,␈α∨however,␈α∨the␈α∨problem␈α∨of
␈↓ α←␈↓inconsistency␈αdetection␈α
would␈αremain␈αdifficult.␈α
 One␈αpossible␈αapproach␈α
would
␈↓ α←␈↓be␈α⊂to␈α⊂view␈α⊂the␈α⊂rule␈α⊂base␈α⊂as␈α⊂a␈α⊂directed␈α⊂graph,␈α⊂where␈α⊂nodes␈α⊂correspond␈α∂to
␈↓ α←␈↓attributes,␈α
and␈α
links␈α∞correspond␈α
to␈α
rules,␈α
with␈α∞weights␈α
on␈α
the␈α
links␈α∞equal␈α
to
␈↓ α←␈↓the␈α
certainty␈α∞factors␈α
of␈α∞the␈α
rules.␈α
 In␈α∞these␈α
terms,␈α∞the␈α
example␈α∞above␈α
would
␈↓ α←␈↓look␈α⊂like␈α⊂two␈α∂different␈α⊂paths␈α⊂from␈α⊂category␈α∂to␈α⊂aerobicity.␈α⊂ We␈α⊂might␈α∂then
␈↓ α←␈↓attempt␈αto␈αdevelop␈αa␈αtaxonomy␈αof␈αtopologic␈αforms␈αthat␈αwere␈α``suspicious''␈αand
␈↓ α←␈↓take␈α
the␈αproper␈α
action␈α(or␈α
just␈α
warn␈αthe␈α
expert)␈αif␈α
a␈α
new␈αrule␈α
ever␈αresulted␈α
in
␈↓ α←␈↓such␈α∀a␈α∀form.␈α∀ (Note␈α∃that␈α∀the␈α∀problem␈α∀is␈α∃a␈α∀good␈α∀deal␈α∀easier␈α∃with␈α∀this
␈↓ α←␈↓incremental␈α≠approach.␈α≠ By␈α≠guaranteeing␈α~the␈α≠integrity␈α≠of␈α≠the␈α~current
␈↓ α←␈↓knowledge base, there is far less work to do when a new rule is added.)
␈↓ α←␈↓␈↓ β?``Semantic''␈α
conflicts␈α∞are␈α
more␈α∞difficult␈α
to␈α∞handle.␈α
 How,␈α∞for␈α
instance,
␈↓ α←␈↓can␈α
the␈αsystem␈α
know␈αthat␈α
it␈αis␈α
a␈α
contradiction␈αto␈α
have␈αtwo␈α
rules␈αconclude␈α
that
␈↓ α←␈↓␈↓↓it␈αis␈αdefinite␈αthat␈α
the␈αsite␈αof␈αthe␈αculture␈α
is␈αone␈αof␈αthe␈α
sterile␈αsites␈↓␈αand␈αthat␈α␈↓↓it␈α
is
␈↓ α←␈↓↓definite␈α
that␈α
the␈α
site␈αis␈α
one␈α
of␈α
the␈α
nonsterile␈αsites␈↓.␈α
 The␈α
important␈α
point␈αis␈α
that

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[20]␈α
This␈α
is␈α
not␈αunreasonable.␈α
 Conditional␈α
probabilities␈α
in␈αstandard␈α
Boolean
␈↓ α←␈↓logic␈αpresent␈αan␈αanalogous␈αsituation.␈α Given␈αan␈αhypothesis␈αH␈αand␈αa␈αpiece␈αof
␈↓ α←␈↓evidence␈α
E,␈α
E␈αcan␈α
be␈α
both␈α
evidence␈αin␈α
favor␈α
of␈α
H␈α(P[H|E]␈α
=␈α
.8)␈αand␈α
evidence
␈↓ α←␈↓in favor of not-H (P[not-H|E] = .2).
␈↓ α←␈↓␈↓122    KNOWLEDGE ACQUISITION I␈↓ 
#5-8␈↓

␈↓"β␈↓ α←␈↓these␈αtwo␈αclasses␈αare␈αmutually␈αexclusive.␈α There␈αmight␈αbe␈αa␈αrule␈αof␈αthe␈αform
␈↓ α←␈↓␈↓↓if␈αa␈αsite␈αis␈αsterile␈αthen␈αit␈αis␈αnot␈αnonsterile␈↓,␈αbut␈αattempting␈αto␈αaccount␈αfor␈αsuch
␈↓ α←␈↓problems␈α
on␈α
a␈α∞case-by-case␈α
basis␈α
is␈α
difficult␈α∞to␈α
do␈α
for␈α
any␈α∞large␈α
knowledge
␈↓ α←␈↓base.␈α
 Consider,␈α
for␈αinstance,␈α
the␈α
difficulty␈αthat␈α
would␈α
be␈α
encountered␈αwhen
␈↓ α←␈↓it␈α∞became␈α∞clear␈α∞that␈α∂the␈α∞two␈α∞classes␈α∞are␈α∞not␈α∂exhaustive␈α∞and␈α∞that␈α∞there␈α∂is␈α∞a
␈↓ α←␈↓third␈α∪class␈α∪of␈α∪sites␈α∪that␈α∪are␈α∪``questionably''␈α∪sterile.␈α∪ A␈α∪more␈α∪fundamental
␈↓ α←␈↓solution␈α⊂to␈α⊃the␈α⊂problem␈α⊂requires,␈α⊃once␈α⊂again,␈α⊂an␈α⊃extensive␈α⊂body␈α⊃of␈α⊂world
␈↓ α←␈↓knowledge␈αnot␈α
currently␈αa␈αpart␈α
of␈αthe␈αsystem.␈α
 We␈αhave␈αnot␈α
yet␈αinvestigated
␈↓ α←␈↓the question of dealing with this sort of conflict.

␈↓ α←␈↓␈↓αLimits of the interactive transfer of expertise approach␈↓    
␈↓ α←␈↓␈↓ β?While␈α∞␈↓¬TEIRESIAS␈↓␈α∞has␈α∞not␈α∞as␈α∞yet␈α∞been␈α∞tested␈α∞by␈α∞having␈α∞experts␈α∞use␈α
it,
␈↓ α←␈↓our␈αexperience␈αwith␈αmanual␈αknowledge␈αacquisition␈αprovides␈αa␈αperspective␈αon
␈↓ α←␈↓its␈αlikely␈αarea␈αof␈αgreatest␈αutility.␈α As␈αnoted,␈αour␈αapproach␈αinvolves␈αknowledge
␈↓ α←␈↓transfer␈αthat␈αis␈αinteractive,␈αthat␈αis␈α
set␈αin␈αthe␈αcontext␈αof␈αa␈αshortcoming␈α
in␈αthe
␈↓ α←␈↓knowledge␈α⊂base,␈α⊃and␈α⊂that␈α⊃transfers␈α⊂a␈α⊃few␈α⊂rules␈α⊃at␈α⊂a␈α⊃time.␈α⊂ Each␈α⊃of␈α⊂these
␈↓ α←␈↓implies certain constraints on the range of applicability of this system.
␈↓ α←␈↓␈↓ β?Interactive␈α∀knowledge␈α∀transfer␈α∀seems␈α∀best␈α∀suited␈α∀to␈α∀task␈α∪domains
␈↓ α←␈↓involving␈αproblem␈αsolving␈αthat␈αis␈αentirely␈αor␈αprimarily␈αa␈αhigh-level␈α
cognitive
␈↓ α←␈↓task,␈α⊃with␈α∩a␈α⊃number␈α⊃of␈α∩distinct,␈α⊃specifiable␈α⊃principles.␈α∩ Medical␈α⊃diagnosis
␈↓ α←␈↓seems␈αan␈α
appropriate␈αdomain,␈αbut␈α
the␈αtechnique␈αwould␈α
not␈αseem␈α
well␈αsuited
␈↓ α←␈↓to␈α∂those␈α∂parts␈α∂of,␈α∂say,␈α∂speech␈α∂understanding␈α∂or␈α∂scene␈α∂recognition␈α⊂in␈α∂which
␈↓ α←␈↓low-level processes appear to play a significant role.
␈↓ α←␈↓␈↓ β?Knowledge␈α∩acquisition␈α∩in␈α∩context␈α∩appears␈α∩to␈α∩offer␈α∩a␈α∩useful␈α⊃guide
␈↓ α←␈↓wherever␈α⊂knowledge␈α⊂of␈α⊂the␈α⊂domain␈α⊂is␈α⊂as␈α⊂yet␈α⊂ill-specified,␈α⊂but␈α⊃the␈α⊂context
␈↓ α←␈↓need␈α∪not␈α∪be␈α∩a␈α∪single␈α∪consultation,␈α∪as␈α∩used␈α∪here.␈α∪ Our␈α∪recent␈α∩experience
␈↓ α←␈↓suggests␈α∩that␈α∩an␈α∪effective␈α∩context␈α∩is␈α∪also␈α∩provided␈α∩by␈α∪examining␈α∩certain
␈↓ α←␈↓subsets␈α∩of␈α∩rules␈α∩in␈α∩the␈α∪knowledge␈α∩base,␈α∩using␈α∩them␈α∩as␈α∩a␈α∪framework␈α∩for
␈↓ α←␈↓specifying new rules.
␈↓ α←␈↓␈↓ β?Finally,␈α∪the␈α∩rule-at-a-time␈α∪approach␈α∩is␈α∪perhaps␈α∩the␈α∪most␈α∩limiting
␈↓ α←␈↓factor.␈α The␈α
example␈αgiven␈α
earlier␈αworks␈αwell,␈α
of␈αcourse,␈α
because␈αthe␈αbug␈α
was
␈↓ α←␈↓manufactured␈α
by␈α
removing␈α
a␈α
single␈αrule.␈α
 In␈α
general,␈α
the␈α
approach␈αseems␈α
well
␈↓ α←␈↓suited␈αto␈αthe␈αlater␈αstages␈αof␈αknowledge␈αbase␈αconstruction,␈αin␈αwhich␈αbugs␈αmay
␈↓ α←␈↓indeed␈α
be␈αcaused␈α
by␈α
the␈αabsence␈α
of␈α
one␈αor␈α
a␈αfew␈α
rules.␈α
 We␈αneed␈α
not␈α
be␈αas
␈↓ α←␈↓lucky␈α∩as␈α⊃the␈α∩present␈α⊃example,␈α∩in␈α∩which␈α⊃one␈α∩rule␈α⊃repairs␈α∩three␈α∩bugs;␈α⊃the
␈↓ α←␈↓approach␈α↔will␈α↔also␈α↔work␈α↔if␈α↔three␈α↔independent␈α↔bugs␈α↔arise␈α↔in␈α↔a␈α⊗given
␈↓ α←␈↓consultation.␈α⊃ But␈α⊃early␈α⊃in␈α⊂knowledge␈α⊃base␈α⊃construction,␈α⊃where␈α⊃large␈α⊂sub-
␈↓ α←␈↓areas␈α
of␈αa␈α
domain␈αare␈α
not␈α
yet␈αspecified,␈α
it␈αappears␈α
more␈αuseful␈α
to␈α
deal␈αwith
␈↓ α←␈↓groups␈α
of␈αrules␈α
or,␈α
more␈αgenerally,␈α
to␈αdeal␈α
with␈α
larger␈αsegments␈α
of␈α
the␈αbasic
␈↓ α←␈↓task, as for example in [Waterman77].
␈↓ α←␈↓␈↓ β?In␈αgeneral,␈αthen,␈αthis␈αapproach␈αseems␈αwell␈αsuited␈αto␈αthe␈αlater␈αstages␈αof
␈↓ α←␈↓knowledge base construction for systems performing high-level tasks.
␈↓ α←␈↓␈↓5-9␈↓ 	βSUMMARY    123␈↓

␈↓"β␈↓ α←␈↓␈↓α5-9    SUMMARY␈↓
␈↓ α←␈↓␈↓ β?This␈α
chapter␈α
has␈α
explored␈α
and␈αillustrated␈α
a␈α
number␈α
of␈α
issues.␈α First,
␈↓ α←␈↓by␈α∪doing␈α∪knowledge␈α∪acquisition␈α∪in␈α∩the␈α∪context␈α∪of␈α∪a␈α∪shortcoming␈α∪in␈α∩the
␈↓ α←␈↓knowledge␈α_base,␈α→␈↓¬TEIRESIAS␈↓'s␈α_acquisition␈α→system␈α_can␈α_build␈α→up␈α_a␈α→set␈α_of
␈↓ α←␈↓expectations about the class of rule it is going to get.
␈↓ α←␈↓␈↓ β?Second,␈α∀in␈α∀the␈α∀set␈α∀of␈α∃rule␈α∀models␈α∀␈↓¬TEIRESIAS␈↓␈α∀has␈α∀a␈α∀model␈α∃of␈α∀the
␈↓ α←␈↓information␈α∂in␈α∂the␈α∂knowledge␈α∞base,␈α∂and␈α∂it␈α∂is␈α∞by␈α∂selecting␈α∂one␈α∂part␈α∂of␈α∞that
␈↓ α←␈↓model (a specific rule model) that it expresses its expectations.
␈↓ α←␈↓␈↓ β?Third,␈α
because␈α
it␈α
has␈α
a␈αmodel␈α
of␈α
the␈α
knowledge␈α
base,␈α
␈↓¬TEIRESIAS␈↓␈αcan␈α
tell
␈↓ α←␈↓whether␈α
some␈αnew␈α
piece␈αof␈α
information␈α``fits␈α
in''␈αto␈α
what␈αis␈α
already␈αknown.␈α
 It
␈↓ α←␈↓is␈α⊂the␈α⊂occurrence␈α⊂of␈α⊂a␈α⊂partial␈α⊂match␈α⊂between␈α⊂the␈α⊂expectations␈α⊂in␈α⊃the␈α⊂rule
␈↓ α←␈↓model␈α
and␈αthe␈α
new␈α
rule␈αthat␈α
prompts␈αthe␈α
system␈α
to␈αmake␈α
suggestions␈α
to␈αthe
␈↓ α←␈↓expert.
␈↓ α←␈↓␈↓ β?Fourth,␈αthe␈αrule␈αmodels␈αare␈αcomputed␈αdirectly␈αfrom␈αthe␈αset␈αof␈αrules␈αin
␈↓ α←␈↓the␈α∞knowledge␈α∞base␈α∞and␈α∞are␈α∞updated␈α∞whenever␈α∞a␈α∞new␈α∞rule␈α∞is␈α∞added.␈α
 This
␈↓ α←␈↓means␈α⊃that␈α⊃the␈α⊃system's␈α⊃model␈α⊃of␈α⊃its␈α⊃knowledge␈α⊃is␈α⊃both␈α⊃derived␈α⊃from␈α⊂its
␈↓ α←␈↓``experience'' and constantly evolving along with the knowledge base itself.
␈↓ α←␈↓␈↓␈↓ 
α    125␈↓




␈↓"β␈↓ α←␈↓␈↓ ε∃␈↓αChapter 6



␈↓"β␈↓ α←␈↓α␈↓ αt␈↓λKNOWLEDGE ACQUISITION II␈↓α


␈↓"β␈↓ α←␈↓α␈↓ ¬5learning new primitives









␈↓"β␈↓ α←␈↓␈↓ ¬
Yes, but I see that even your own words miss the mark.... 
␈↓"β␈↓ α←␈↓␈↓ πs␈↓↓Oedipus the King␈↓, line 324

␈↓"β␈↓ α←␈↓␈↓α6-1    INTRODUCTION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
techniques␈α
described␈αin␈α
the␈α
previous␈αchapter␈α
make␈α
it␈αpossible␈α
for
␈↓ α←␈↓the␈α∪expert␈α∪to␈α∪teach␈α∩the␈α∪system␈α∪new␈α∪rules,␈α∩expressed␈α∪in␈α∪terms␈α∪of␈α∩known
␈↓ α←␈↓concepts.␈α But␈αthis␈α
capability␈αalone␈αwould␈α
be␈αinsufficient␈αfor␈α
any␈αsubstantial
␈↓ α←␈↓education␈αof␈αthe␈αsystem␈αsince␈αgaps␈α
in␈αthe␈αknowledge␈αbase␈αmight␈αrequire␈α
rules
␈↓ α←␈↓dealing␈α∞with␈α∞concepts␈α∞not␈α∞yet␈α∞known␈α∞to␈α∞the␈α∞system.␈α∞ This␈α∞chapter␈α∞describes
␈↓ α←␈↓how␈α
the␈α
expert␈α
can␈α
teach␈α
the␈α
system␈α
new␈α
conceptual␈α
primitives␈α
and␈α
new␈α
types
␈↓ α←␈↓of conceptual primitives.␈↓
1␈↓
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∀capability␈α∀requires␈α∪dealing␈α∀with␈α∀a␈α∪new␈α∀set␈α∀of␈α∀problems,␈α∪in
␈↓ α←␈↓addition␈α∀to␈α∀those␈α∀faced␈α∀earlier.␈α∀ There␈α∀will,␈α∀in␈α∀particular,␈α∀be␈α∃a␈α∀greater
␈↓ α←␈↓emphasis␈α∩on␈α∩the␈α∩manipulation␈α∩of␈α⊃data␈α∩structures␈α∩in␈α∩the␈α∩knowledge␈α⊃base.
␈↓ α←␈↓Acquisition␈αof␈αnew␈αrules␈αdealt␈αwith␈αa␈αsingle␈αtype␈αof␈αstructure,␈αone␈αwhich␈αwas
␈↓ α←␈↓understood␈α⊂in␈α∂terms␈α⊂of␈α∂a␈α⊂combination␈α∂of␈α⊂available␈α∂primitives.␈α⊂ There␈α∂was
␈↓ α←␈↓thus␈α∀a␈α∀single,␈α∀uniform␈α∀process␈α∀for␈α∀acquisition␈α∀and␈α∀integration,␈α∃with␈α∀an
␈↓ α←␈↓emphasis␈α
on␈α∞understanding␈α
and␈α∞interpreting␈α
the␈α
English␈α∞text.␈α
 Here,␈α∞in␈α
the
␈↓ α←␈↓acquisition␈α
of␈α
new␈α
conceptual␈α
primitives,␈α
it␈α
is␈α
necessary␈α
to␈α
deal␈α
with␈α
a␈αwide
␈↓ α←␈↓range␈α
of␈α∞data␈α
structures,␈α
each␈α∞of␈α
which␈α
may␈α∞have␈α
its␈α
own␈α∞requirements␈α
for

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[1]␈α⊂A␈α⊂``new␈α⊂conceptual␈α⊂primitive''␈α⊂means␈α∂a␈α⊂new␈α⊂instance␈α⊂of␈α⊂one␈α⊂of␈α⊂the␈α∂13
␈↓ α←␈↓primitives␈α⊃listed␈α∩in␈α⊃Section␈α∩2-4-4.␈α⊃ A␈α⊃``new␈α∩type␈α⊃of␈α∩conceptual␈α⊃primitive''
␈↓ α←␈↓refers␈α
to␈α
teaching␈α
the␈α
system␈α
about␈α
a␈α
new␈α
kind␈α
of␈α
primitive␈α
in␈α
addition␈αto␈α
the
␈↓ α←␈↓existing 13.
␈↓ α←␈↓␈↓126    KNOWLEDGE ACQUISITION II␈↓ 
#6-1␈↓

␈↓"β␈↓ α←␈↓integration␈α∪into␈α∪the␈α∪knowledge␈α∪base.␈α∪ The␈α∪problem␈α∪thus␈α∪has␈α∪two␈α∪major
␈↓ α←␈↓aspects␈α
to␈αit:␈α
(␈↓↓i␈↓)␈α
knowledge␈αacquisition␈α
and␈α(␈↓↓ii␈↓)␈α
knowledge␈α
base␈αmanagement.
␈↓ α←␈↓In␈α∂response,␈α∂the␈α∞techniques␈α∂used␈α∂address␈α∞both␈α∂the␈α∂difficulties␈α∂presented␈α∞by
␈↓ α←␈↓the␈α⊂knowledge␈α⊂transfer␈α⊂process␈α⊂and␈α∂the␈α⊂general␈α⊂issues␈α⊂of␈α⊂constructing␈α∂and
␈↓ α←␈↓maintaining a large collection of data structures.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α⊗chapter␈α⊗is␈α∃divided␈α⊗into␈α⊗three␈α∃main␈α⊗parts.␈α⊗ The␈α⊗first␈α∃part
␈↓ α←␈↓introduces␈α∪the␈α∪idea␈α∀of␈α∪a␈α∪␈↓↓data␈α∪structure␈α∀schema␈↓,␈α∪a␈α∪device␈α∀for␈α∪describing
␈↓ α←␈↓representations,␈α∞and␈α∞contains␈α∞the␈α∞bulk␈α∞of␈α∞the␈α∞discussion␈α∞about␈α∞it.␈α∞ It␈α
begins
␈↓ α←␈↓with␈αa␈αgeneral␈αoverview␈αof␈αthe␈αfundamental␈αproblems␈αattacked␈αand␈αthe␈αbasic
␈↓ α←␈↓ideas␈αused.␈α It␈αthen␈αcontinues␈αwith␈αtraces␈αthat␈αshow␈αhow␈α␈↓¬TEIRESIAS␈↓␈αdirects␈αthe
␈↓ α←␈↓acquisition␈αof␈αa␈αnew␈αvalue␈αfor␈αan␈αattribute,␈αdemonstrating␈αa␈α
simple␈αexample
␈↓ α←␈↓of␈α∞``filling␈α∞out''␈α∞an␈α∞existing␈α∞schema␈α∞to␈α∞produce␈α∞a␈α∞new␈α∞instance.␈α∂ Section␈α∞6-8
␈↓ α←␈↓and␈α⊃Section␈α⊃6-9␈α⊃then␈α⊃discuss␈α⊃the␈α⊃organization␈α⊃and␈α⊃use␈α⊃of␈α⊃the␈α⊂knowledge
␈↓ α←␈↓carried in the schemata.
␈↓"β␈↓ α←␈↓␈↓ β?Section␈α_6-10␈α_starts␈α_the␈α_second␈α_part␈α_with␈α_an␈α_example␈α_and␈α↔an
␈↓ α←␈↓explanation␈α
of␈α
the␈α
acquisition␈α
of␈αa␈α
new␈α
attribute.␈α
 This␈α
part␈αdemonstrates␈α
the
␈↓ α←␈↓use␈α∞of␈α∂the␈α∞schemata␈α∂on␈α∞more␈α∞complex␈α∂data␈α∞structures␈α∂and␈α∞indicates␈α∂how␈α∞a
␈↓ α←␈↓new␈α∃schema␈α∃can␈α∃be␈α∃acquired␈α∃using␈α∃the␈α∃same␈α∃techniques␈α⊗employed␈α∃for
␈↓ α←␈↓acquiring new instances.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
last␈αpart␈α
begins␈α
with␈αSection␈α
6-12,␈α
which␈αdescribes␈α
how␈α
to␈αstart
␈↓ α←␈↓the␈αknowledge␈α
acquisition␈αprocess␈αwhen␈α
building␈αan␈αentirely␈α
new␈αknowledge
␈↓ α←␈↓base.  It shows an example of ␈↓¬TEIRESIAS␈↓'s performance on this task.

␈↓"β␈↓ α←␈↓␈↓α6-2    KEY IDEAS:   OVERVIEW␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂discussion␈α∂below␈α∂requires␈α∂including␈α∂a␈α∂certain␈α∂amount␈α∂of␈α∞detail
␈↓ α←␈↓concerning␈αboth␈αthe␈αschemata␈αand␈αinternal␈αdata␈αstructure␈αimplementation.␈α It
␈↓ α←␈↓also␈αranges␈αover␈αa␈αlarge␈αnumber␈αof␈αtopics␈αand␈αexamines␈αsteps␈αtoward␈αsolving
␈↓ α←␈↓many␈αof␈αthe␈αproblems.␈α To␈αinsure␈αthat␈αthe␈αmore␈αimportant␈αideas␈αare␈αnot␈αlost
␈↓ α←␈↓in␈α∩the␈α⊃mass␈α∩of␈α⊃detail,␈α∩they␈α⊃are␈α∩summarized␈α⊃below␈α∩and␈α⊃labelled␈α∩with␈α⊃the
␈↓ α←␈↓section in which they first appear.
␈↓"β␈↓ α←␈↓␈↓ β?The most basic observation we make is that

␈↓"β␈↓ α←␈↓␈↓ ββ(1)␈↓ β?By␈α
supplying␈α
a␈α
system␈α
with␈α
a␈α
store␈α
of␈α
knowledge␈α
about␈α
its␈α
own
␈↓ α←␈↓␈↓ β?representations,␈α∀both␈α∀knowledge␈α∀acquisition␈α∀and␈α∪knowledge
␈↓ α←␈↓␈↓ β?base␈α⊂management␈α∂can␈α⊂be␈α∂carried␈α⊂out␈α∂in␈α⊂a␈α⊂high-level␈α∂dialog
␈↓ α←␈↓␈↓ β?that transfers information relatively easily.  Section 6-5


␈↓ α←␈↓Further␈α!observations␈α!dealing␈α!with␈α!the␈α!store␈α!of␈α!knowledge␈α about
␈↓ α←␈↓representations include:

␈↓"β␈↓ α←␈↓␈↓ ββ(2)␈↓ β?Each␈αof␈α
the␈αconceptual␈αprimitives␈α
(knowledge␈αrepresentations)
␈↓ α←␈↓␈↓ β?from␈α∀which␈α∀rules␈α∀(and␈α∃other␈α∀structures)␈α∀are␈α∀built␈α∃will␈α∀be
␈↓ α←␈↓␈↓ β?viewed␈α∂as␈α∂an␈α∞extended␈α∂data␈α∂type.␈α∞ Each␈α∂such␈α∂extended␈α∞data
␈↓ α←␈↓␈↓ β?type␈α⊃is␈α∩described␈α⊃by␈α∩a␈α⊃␈↓↓data␈α∩structure␈α⊃schema␈↓,␈α∩a␈α⊃record-like
␈↓ α←␈↓␈↓6-2␈↓ πXKEY IDEAS:   OVERVIEW    127␈↓

␈↓"β␈↓ α←␈↓␈↓ β?structure␈αaugmented␈α
with␈αadditional␈α
information␈α(such␈αas␈α
data
␈↓ α←␈↓␈↓ β?structure interrelations).  Section 6-5

␈↓"β␈↓ α←␈↓␈↓ ββ(3)␈↓ β?The␈α~schemata␈α≠provide␈α~a␈α≠language␈α~and␈α≠mechanism␈α~for
␈↓ α←␈↓␈↓ β?describing␈α
representations␈α
and␈α
hence␈α
offer␈α
a␈α
way␈αof␈α
expressing
␈↓ α←␈↓␈↓ β?a body of knowledge about them.  Section 6-6-2

␈↓"β␈↓ α←␈↓␈↓ ββ(4)␈↓ β?The␈α%body␈α%of␈α%knowledge␈α%is␈α%organized␈α%around␈α%the
␈↓ α←␈↓␈↓ β?representational␈α⊂primitives␈α⊂in␈α⊂use␈α⊂(such␈α⊂as,␈α⊂in␈α⊂this␈α⊃case,␈α⊂the
␈↓ α←␈↓␈↓ β?notion of ␈↓↓attribute␈↓, ␈↓↓object␈↓, ␈↓↓value␈↓, etc.).  Section 6-8-2

␈↓"β␈↓ α←␈↓␈↓ ββ(5)␈↓ β?Knowledge␈α⊃is␈α∩represented␈α⊃as␈α⊃a␈α∩collection␈α⊃of␈α∩prototypes␈α⊃(the
␈↓ α←␈↓␈↓ β?schemata).  Section 6-8

␈↓"β␈↓ α←␈↓␈↓ ββ(6)␈↓ β?Knowledge␈α∀can␈α∀be␈α∀viewed␈α∀in␈α∀terms␈α∀of␈α∀different␈α∀levels␈α∪of
␈↓ α←␈↓␈↓ β?generality:␈α1(␈↓↓i␈↓)  schema␈α1instances,␈α2(␈↓↓ii␈↓)  schemata,␈α1and
␈↓ α←␈↓␈↓ β?(␈↓↓iii␈↓)  ``schema-schema.''  Section 6-13

␈↓"β␈↓ α←␈↓␈↓ ββ(7)␈↓ β?The␈α
techniques␈α
we␈α
use␈α
gain␈αa␈α
certain␈α
degree␈α
of␈α
generality␈αby
␈↓ α←␈↓␈↓ β?keeping␈α∂the␈α∞knowledge␈α∂carefully␈α∞stratified␈α∂according␈α∂to␈α∞those
␈↓ α←␈↓␈↓ β?levels.  Section 6-13

␈↓"β␈↓ α←␈↓␈↓ ββ(8)␈↓ β?The␈α≥set␈α≥of␈α≥schemata␈α≥can␈α≥itself␈α≥be␈α≥organized␈α≥into␈α≤a
␈↓ α←␈↓␈↓ β?generalization hierarchy.  Section 6-8-1.


␈↓ α←␈↓Observations dealing with knowledge base management include:

␈↓"β␈↓ α←␈↓␈↓ ββ(9)␈↓ β?It␈αis␈αuseful␈αto␈αconsider␈αthe␈αterms␈α␈↓↓data␈αstructure␈↓,␈α␈↓↓extended␈αdata
␈↓ α←␈↓↓␈↓ β?type␈↓,␈α
and␈α
␈↓↓knowledge␈α
representation␈↓␈α
as␈αinterchangeable.␈α
 Section
␈↓ α←␈↓␈↓ β?6-6-1

␈↓"β␈↓ α←␈↓␈↓ ββ(10)␈↓ β?The␈αsystem␈αis␈α
``totally␈αtyped''␈αin␈α
the␈αsense␈αthat␈α
ideas␈α(4)␈αand␈α
(5)
␈↓ α←␈↓␈↓ β?above␈αare␈αapplied␈αexhaustively␈αto␈αall␈αrepresentations␈αand␈αdata
␈↓ α←␈↓␈↓ β?structures in the system.  Section 6-6-3.

␈↓"β␈↓ α←␈↓␈↓ ββ(11)␈↓ β?Unlike␈α
ordinary␈α
record␈α
structures␈α
or␈α
declarations,␈αthe␈α
schemata
␈↓ α←␈↓␈↓ β?are␈αa␈αpart␈α
of␈αthe␈αsystem␈αitself␈α
and␈αare␈αavailable␈αto␈α
the␈αsystem
␈↓ α←␈↓␈↓ β?for examination.  Section 6-6-3


␈↓ α←␈↓Ideas relevant to knowledge acquisition include the suggestions that:

␈↓"β␈↓ α←␈↓␈↓ ββ(12)␈↓ β?Knowledge␈α≤acquisition␈α≤can␈α≠proceed␈α≤by␈α≤interpreting␈α≠the
␈↓ α←␈↓␈↓ β?information␈α∂in␈α⊂the␈α∂schemata␈α⊂as␈α∂a␈α∂set␈α⊂of␈α∂instructions␈α⊂for␈α∂the
␈↓ α←␈↓␈↓128    KNOWLEDGE ACQUISITION II␈↓ 
#6-2␈↓

␈↓"β␈↓ α←␈↓␈↓ β?construction␈α_and␈α→maintenance␈α_of␈α_the␈α→relevant␈α_knowledge
␈↓ α←␈↓␈↓ β?representations.␈α" Hence␈α"it␈α"is␈α"the␈α"process␈α#of␈α"schema
␈↓ α←␈↓␈↓ β?instantiation that drives knowledge acquisition.  Section 6-9

␈↓"β␈↓ α←␈↓␈↓ ββ(13)␈↓ β?Doing␈αknowledge␈αacquisition␈αvia␈αthe␈αschemata␈αoffers␈αa␈αcertain
␈↓ α←␈↓␈↓ β?level of knowledge base integrity.  Section 6-6-4

␈↓"β␈↓ α←␈↓␈↓ ββ(14)␈↓ β?Acquisition␈α⊗of␈α↔a␈α⊗new␈α⊗␈↓↓instance␈↓␈α↔of␈α⊗an␈α↔existing␈α⊗conceptual
␈↓ α←␈↓␈↓ β?primitive␈α⊃is␈α⊃structured␈α⊃as␈α⊃a␈α⊃process␈α⊃of␈α⊃descent␈α∩through␈α⊃the
␈↓ α←␈↓␈↓ β?schema hierarchy noted above.  Section 6-8-1

␈↓"β␈↓ α←␈↓␈↓ ββ(15)␈↓ β?Acquisition␈α≠of␈α≠a␈α≠new␈α≠␈↓↓kind␈↓␈α≠of␈α≠conceptual␈α≤primitive␈α≠is
␈↓ α←␈↓␈↓ β?structured␈αas␈αa␈αprocess␈αof␈αadding␈αnew␈αbranches␈αto␈αthe␈αschema
␈↓ α←␈↓␈↓ β?hierarchy.  Section 6-10


␈↓"β␈↓ α←␈↓␈↓α6-3    THE FUNDAMENTAL PROBLEM␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Viewed␈α→from␈α→the␈α~perspective␈α→of␈α→knowledge␈α~representation␈α→and
␈↓ α←␈↓knowledge␈α⊃base␈α⊃management,␈α⊃the␈α⊃problem␈α⊃of␈α⊃acquiring␈α⊃a␈α⊃new␈α⊃conceptual
␈↓ α←␈↓primitive␈αcan␈αbe␈αseen␈α
in␈αterms␈αof␈αadding␈αa␈α
new␈αinstance␈αof␈αan␈αextended␈α
data
␈↓ α←␈↓type␈α∀to␈α∀a␈α∪large␈α∀program.␈α∀ Using␈α∪the␈α∀standard␈α∀approach,␈α∀a␈α∪programmer
␈↓ α←␈↓attempting␈α∩this␈α⊃task␈α∩would␈α∩have␈α⊃to␈α∩gather␈α∩a␈α⊃wide␈α∩range␈α∩of␈α⊃information,
␈↓ α←␈↓including␈αthe␈αstructure␈αof␈αthe␈αdata␈αtype␈αand␈αits␈αinterrelations␈αwith␈αother␈αdata
␈↓ α←␈↓types␈α
in␈α
the␈α
program.␈α
 Such␈αinformation␈α
is␈α
typically␈α
recorded␈α
informally␈α(if␈α
at
␈↓ α←␈↓all)␈α
and␈α
is␈α
often␈α
scattered␈α
through␈α
a␈α
range␈α
of␈α
sources;␈α
it␈α
might␈α
be␈α∞found␈α
in
␈↓ α←␈↓comments␈α≤in␈α≤program␈α≤code,␈α≠in␈α≤documents␈α≤and␈α≤manuals␈α≠maintained
␈↓ α←␈↓separately,␈αand␈αin␈αthe␈αmind␈αof␈α
the␈αprogram␈αarchitect.␈α Just␈αfinding␈αall␈αof␈α
this
␈↓ α←␈↓information␈αcan␈αbe␈αa␈αmajor␈αtask,␈αespecially␈αfor␈αsomeone␈αunfamiliar␈αwith␈αthe
␈↓ α←␈↓program.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α
this␈α
situation,␈α
two␈α
sorts␈α∞of␈α
errors␈α
are␈α
common: ␈α
The␈α∞new␈α
instance
␈↓ α←␈↓may␈α
be␈α
given␈αthe␈α
wrong␈α
structure␈αor␈α
it␈α
may␈αbe␈α
improperly␈α
integrated␈αinto␈α
the
␈↓ α←␈↓rest␈α⊃of␈α∩the␈α⊃program.␈α⊃ Since␈α∩an␈α⊃extended␈α∩data␈α⊃type␈α⊃may␈α∩be␈α⊃built␈α∩from␈α⊃a
␈↓ α←␈↓complex␈α∂collection␈α∂of␈α∂components␈α∂and␈α∂pointers,␈α∂it␈α∂is␈α∂not␈α∂uncommon␈α∂that␈α∞a
␈↓ α←␈↓new␈α∪instance␈α∪receives␈α∀an␈α∪incorrect␈α∪internal␈α∪organization,␈α∀that␈α∪extraneous
␈↓ α←␈↓structures␈α
are␈αincluded,␈α
or␈αthat␈α
necessary␈αelements␈α
are␈α
inadvertently␈αomitted.
␈↓ α←␈↓Since␈α
data␈α
structures␈α
in␈α
a␈α
program␈α
are␈α
not␈α
typically␈α
independent,␈α
the␈α
addition
␈↓ α←␈↓of␈α∂a␈α∞new␈α∂instance␈α∞often␈α∂requires␈α∞significant␈α∂effort␈α∞to␈α∂maintain␈α∂the␈α∞existing
␈↓ α←␈↓interdependencies.␈α Errors␈αcan␈αresult␈αfrom␈αdoing␈αthis␈αincorrectly,␈αby␈αviolating
␈↓ α←␈↓the␈α∀interrelationships␈α∀of␈α∀existing␈α∃structures␈α∀or␈α∀(as␈α∀is␈α∀more␈α∃common)␈α∀by
␈↓ α←␈↓omitting a necessary bookkeeping step.

␈↓"β␈↓ α←␈↓␈↓α6-4    SOURCES OF DIFFICULTY␈↓
␈↓"β␈↓ α←␈↓␈↓ β?A␈α
basic␈α
source␈α
of␈α
difficulty␈αin␈α
solving␈α
the␈α
problem␈α
of␈α
acquiring␈αa␈α
new
␈↓ α←␈↓conceptual␈α⊃primitive␈α⊃arises␈α⊃from␈α⊃our␈α⊂desire␈α⊃to␈α⊃deal␈α⊃with␈α⊃the␈α⊃issues␈α⊂noted
␈↓ α←␈↓above␈α≠(insuring␈α≠correct␈α≠structure␈α≤for␈α≠a␈α≠newly␈α≠added␈α≤primitive␈α≠and
␈↓ α←␈↓␈↓6-4␈↓ πTSOURCES OF DIFFICULTY    129␈↓

␈↓"β␈↓ α←␈↓maintaining␈αexisting␈αinterrelationships)␈αin␈αthe␈αcontext␈αof␈αthe␈αglobal␈αgoals␈αset
␈↓ α←␈↓out␈αat␈αthe␈αbeginning.␈α That␈αis,␈αa␈αnonprogrammer␈αshould␈αbe␈αable␈αto␈αbuild␈αthe
␈↓ α←␈↓knowledge␈α
base␈α∞and␈α
be␈α
able␈α∞to␈α
assemble␈α
large␈α∞amounts␈α
of␈α∞knowledge.␈α
 The
␈↓ α←␈↓first␈αof␈α
these␈αgoals␈αmeans␈α
that␈αthe␈αuser␈α
cannot␈α(nor␈αshould␈α
he␈αbe␈αexpected␈α
to)
␈↓ α←␈↓deal␈α∞with␈α∞the␈α∞system␈α∞at␈α
the␈α∞level␈α∞of␈α∞data␈α∞structures;␈α
he␈α∞needs␈α∞a␈α∞dialog␈α∞at␈α
a
␈↓ α←␈↓higher␈α∩level.␈α∩ This␈α⊃is␈α∩accomplished␈α∩by␈α∩having␈α⊃␈↓¬TEIRESIAS␈↓␈α∩take␈α∩care␈α∩of␈α⊃the
␈↓ α←␈↓``details'' and having the user supply only domain-specific information.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩second␈α∩global␈α∩goal--building␈α∩a␈α∩large␈α∪knowledge␈α∩base--brings
␈↓ α←␈↓with␈α
it␈α
the␈α
problem␈α
of␈α
complexity.␈α
 There␈α
is␈α
a␈α
well-known␈α
phenomenon␈α
in␈α
all
␈↓ α←␈↓programs␈α(but␈αmost␈αobvious␈α
in␈αlarge␈αsystems)␈αreferred␈α
to␈αas␈αthe␈α``1␈α
+␈αepsilon
␈↓ α←␈↓bug''␈α
phenomenon:␈αA␈α
change␈αintroduced␈α
to␈α
fix␈αa␈α
known␈αbug␈α
may␈α
result,␈αon
␈↓ α←␈↓the␈αaverage,␈α
in␈αthe␈α
creation␈αof␈α
␈↓↓more␈αthan␈↓␈α
one␈αnew␈α
bug.␈α The␈α
system␈αmay␈α
thus
␈↓ α←␈↓be␈αinherently␈αunstable,␈αsince␈αany␈α
attempt␈αto␈αrepair␈αa␈αproblem␈α
may␈αintroduce
␈↓ α←␈↓more problems than it repairs.
␈↓"β␈↓ α←␈↓␈↓ β?Complexity␈α
arises␈αin␈α
our␈α
case␈αprimarily␈α
because␈α
of␈αsize:␈α
the␈α
size␈αof␈α
the
␈↓ α←␈↓performance␈αprogram's␈αknowledge␈αbase␈αand␈αthe␈αwealth␈αof␈αdetail␈αinvolved␈αin
␈↓ α←␈↓fully␈αdescribing␈αits␈αknowledge␈αrepresentations.␈α There␈αis,␈αfor␈αexample,␈αa␈αlarge
␈↓ α←␈↓number␈αof␈αdifferent␈αdata␈αtypes,␈αeach␈αwith␈αits␈αown␈αstructural␈αorganization,␈αits
␈↓ α←␈↓own␈α_set␈α_of␈α_interrelations␈α_with␈α_other␈α_data␈α_types,␈α_and␈α_its␈α_own␈α_set␈α↔of
␈↓ α←␈↓requirements␈αfor␈αintegration␈αinto␈αthe␈αprogram.␈α There␈αis␈αalso␈αa␈αlarge␈αnumber
␈↓ α←␈↓of␈α⊂instances␈α⊂of␈α∂each␈α⊂data␈α⊂type.␈α∂ Since␈α⊂modifications␈α⊂to␈α∂a␈α⊂data␈α⊂type␈α∂design
␈↓ α←␈↓have␈α⊃to␈α⊃be␈α⊃carried␈α⊃out␈α⊃on␈α⊃all␈α⊃of␈α⊃its␈α⊃instances,␈α⊃the␈α⊃efficient␈α⊃retrieval␈α⊃and
␈↓ α←␈↓processing␈α∞of␈α∂this␈α∞set␈α∂is␈α∞another␈α∂problem␈α∞that␈α∂involves␈α∞the␈α∂management␈α∞of
␈↓ α←␈↓large numbers of structures.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α
order␈α
to␈α
make␈α
the␈α
acquisition␈α
of␈α
new␈α
conceptual␈α
primitives␈α
possible
␈↓ α←␈↓in␈αthe␈αcontext␈α
of␈αour␈αoriginal␈αgoals,␈α
then,␈αwe␈αhave␈αto␈α
provide␈αthe␈αuser␈αwith␈α
a
␈↓ α←␈↓system␈α∪that␈α∀carries␈α∪on␈α∪a␈α∀high-level␈α∪dialog␈α∪and␈α∀that␈α∪keeps␈α∪track␈α∀of␈α∪the
␈↓ α←␈↓numerous␈αdetails␈αof␈αdata␈αstructure␈αimplementation.␈α The␈αfirst␈αof␈αthese␈αdesign
␈↓ α←␈↓requirements␈α∂will␈α∞insure␈α∂that␈α∂the␈α∞system␈α∂is␈α∂comprehensible␈α∞to␈α∂the␈α∂user;␈α∞the
␈↓ α←␈↓second␈α
will␈α∞insure␈α
that␈α
new␈α∞bugs␈α
are␈α
not␈α∞inadvertently␈α
created␈α∞while␈α
fixing
␈↓ α←␈↓old ones.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈αthat␈αthis␈αemphasis␈αon␈αthe␈αnecessity␈αof␈αavoiding␈αbugs␈αduring␈αthe
␈↓ α←␈↓process␈α∩of␈α⊃acquiring␈α∩new␈α∩primitives␈α⊃is␈α∩really␈α∩no␈α⊃different␈α∩than␈α∩the␈α⊃view
␈↓ α←␈↓presented␈α∀earlier␈α∃in␈α∀discussing␈α∀explanation␈α∃and␈α∀rule␈α∃acquisition.␈α∀ When
␈↓ α←␈↓dealing␈α
with␈αrules,␈α
we␈αnoted␈α
that␈α
the␈αlarge␈α
amount␈αof␈α
knowledge␈αrequired␈α
for
␈↓ α←␈↓high␈αperformance␈αmakes␈αshortcomings␈αin␈αthe␈αknowledge␈αbase␈αinevitable,␈αand
␈↓ α←␈↓we␈α∪emphasized␈α∪the␈α∀benefits␈α∪of␈α∪using␈α∀these␈α∪shortcomings␈α∪to␈α∀provide␈α∪the
␈↓ α←␈↓context␈α~and␈α~focus␈α→for␈α~knowledge␈α~acquisition.␈α→ We␈α~will␈α~similarly␈α→use
␈↓ α←␈↓shortcomings␈α→in␈α→the␈α→knowledge␈α→base␈α→to␈α→provide␈α→the␈α→context␈α→for␈α_the
␈↓ α←␈↓acquisition of new conceptual primitives.
␈↓"β␈↓ α←␈↓␈↓ β?In␈αboth␈αcases,␈αwe␈αmust␈αavoid␈αintroducing␈αerrors␈αduring␈αthe␈αprocess␈αof
␈↓ α←␈↓knowledge␈αacquisition.␈α This␈αis␈αof␈αgreater␈αconcern␈αduring␈αacquisition␈αof␈αnew
␈↓ α←␈↓conceptual␈α⊃primitives␈α∩for␈α⊃reasons␈α∩arising␈α⊃out␈α∩of␈α⊃the␈α∩nature␈α⊃of␈α∩the␈α⊃errors
␈↓ α←␈↓encountered and the objects involved.
␈↓"β␈↓ α←␈↓␈↓ β?As␈αnoted␈αearlier,␈α
conceptual␈αprimitives␈αare␈α
each␈αindividually␈αfar␈α
more
␈↓ α←␈↓␈↓130    KNOWLEDGE ACQUISITION II␈↓ 
#6-4␈↓

␈↓"β␈↓ α←␈↓complex␈αin␈αtheir␈αstructure␈αthan␈αrules.␈α In␈αaddition,␈αthere␈αis␈αonly␈αa␈αsingle␈αrule
␈↓ α←␈↓format,␈α⊗while␈α⊗there␈α⊗are␈α⊗many␈α⊗different␈α⊗conceptual␈α↔primitive␈α⊗structures.
␈↓ α←␈↓There is, thus, far greater opportunity for error.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α⊗addition,␈α↔where␈α⊗the␈α⊗rules␈α↔are␈α⊗designed␈α⊗to␈α↔be␈α⊗fundamentally
␈↓ α←␈↓independent,␈α
data␈α∞structures␈α
used␈α
to␈α∞represent␈α
the␈α
conceptual␈α∞primitives␈α
are
␈↓ α←␈↓often␈αinterrelated␈αin␈αsubtle␈αways␈α
and␈αthe␈αcharacter␈αof␈αthe␈αerrors␈α
produced␈αis
␈↓ α←␈↓very␈α∀different.␈α∀ The␈α∀independence␈α∀of␈α∀rules␈α∀means␈α∀that␈α∀their␈α∪interaction
␈↓ α←␈↓during␈α⊂a␈α∂consultation␈α⊂can␈α∂be␈α⊂understood␈α∂by␈α⊂a␈α∂simple␈α⊂model␈α∂in␈α⊂which␈α∂the
␈↓ α←␈↓contribution␈α∪of␈α∀each␈α∪rule␈α∀is␈α∪considered␈α∀individually.␈α∪ This␈α∀is␈α∪manifestly
␈↓ α←␈↓untrue␈αof␈α
complex␈αdata␈α
structures,␈αwhere␈αerrors␈α
in␈αformat␈α
can␈αresult␈αin␈α
subtle
␈↓ α←␈↓interactions.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈αthere␈αis␈αthe␈αissue␈α
of␈αthe␈αconceptual␈αlevel␈αof␈αthe␈α
objects␈αbeing
␈↓ α←␈↓manipulated,␈α
that␈α
is,␈α
their␈α
likely␈α
familiarity␈α
to␈α
the␈α
expert.␈α
 It␈αseems␈α
reasonable
␈↓ α←␈↓to␈α⊂assume␈α⊂that␈α⊂domain-specific␈α⊂rules␈α⊂will␈α⊂deal␈α⊂with␈α⊂knowledge␈α⊂sufficiently
␈↓ α←␈↓familiar␈α
to␈α
the␈α
expert␈α
that␈α
he␈α
will␈α
be␈α
able␈α
to␈α
understand␈α
the␈α
program␈α
in␈α
these
␈↓ α←␈↓terms.␈α
 We␈α
do␈α
not␈α
assume␈α
that␈α
he␈α
is␈α
familiar␈α
enough␈α
with␈α
data␈α
structures␈α
and
␈↓ α←␈↓representations to be able to manipulate or debug them.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α⊗summary,␈α⊗recall␈α⊗the␈α⊗distinction␈α⊗drawn␈α⊗in␈α⊗chapter␈α⊗1␈α∃between
␈↓ α←␈↓expertise␈α∪and␈α∪formalism.␈α∩ Because␈α∪rules␈α∪are␈α∩designed␈α∪to␈α∪have␈α∪a␈α∩sharply
␈↓ α←␈↓constrained␈α⊂degree␈α⊂of␈α⊂interaction␈α⊃and␈α⊂to␈α⊂be␈α⊂comprehensible␈α⊂to␈α⊃the␈α⊂expert,
␈↓ α←␈↓errors␈α⊂in␈α∂the␈α⊂knowledge␈α∂base␈α⊂may␈α∂properly␈α⊂be␈α∂considered␈α⊂shortcomings␈α∂of
␈↓ α←␈↓expertise.␈α⊂ The␈α⊂complex␈α⊂interrelations␈α⊂of␈α⊂data␈α⊂structures␈α⊂used␈α⊃to␈α⊂represent
␈↓ α←␈↓conceptual␈α∞primitives␈α∂and␈α∞the␈α∂subtle␈α∞nature␈α∂of␈α∞the␈α∂bugs␈α∞they␈α∂produce␈α∞puts
␈↓ α←␈↓them␈αin␈α
the␈αdomain␈α
of␈αerrors␈α
of␈αformalism.␈α
 In␈αdealing␈α
with␈αthe␈α
creation␈αof
␈↓ α←␈↓new␈αconceptual␈αprimitives,␈αtherefore,␈α
strong␈αemphasis␈αis␈αplaced␈αon␈α
techniques
␈↓ α←␈↓that assure a high degree of integrity.

␈↓"β␈↓ α←␈↓␈↓α6-5    THE SOLUTION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?In␈α
the␈α
simplest␈α
terms,␈αthe␈α
solution␈α
we␈α
suggest␈αis␈α
to␈α
give␈α
the␈α
system␈αa
␈↓ α←␈↓store␈α∂of␈α∂knowledge␈α∂about␈α∂its␈α∂representations␈α∂and␈α∂the␈α∂capability␈α∂to␈α⊂use␈α∂this
␈↓ α←␈↓knowledge as a basis for the construction and management of them.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α∀more␈α∀detail: ␈α∪We␈α∀view␈α∀every␈α∪knowledge␈α∀representation␈α∀in␈α∪the
␈↓ α←␈↓system␈αas␈αan␈αextended␈αdata␈αtype.␈α Explicit␈αdescriptions␈αof␈αeach␈αdata␈αtype␈αare
␈↓ α←␈↓written,␈α∩descriptions␈α⊃that␈α∩include␈α⊃all␈α∩the␈α⊃information␈α∩about␈α∩structure␈α⊃and
␈↓ α←␈↓interrelations␈α
that␈αwas␈α
noted␈α
earlier␈αas␈α
often␈α
being␈αwidely␈α
scattered.␈α Next,␈α
we
␈↓ α←␈↓devise␈α
a␈α
language␈α
in␈α∞which␈α
all␈α
of␈α
this␈α∞information␈α
can␈α
be␈α
put␈α∞in␈α
machine-
␈↓ α←␈↓comprehensible␈α
terms␈αand␈α
write␈αthe␈α
descriptions␈αin␈α
those␈αterms,␈α
making␈αthis
␈↓ α←␈↓store␈αof␈αinformation␈αavailable␈αto␈αthe␈αsystem.␈αFinally,␈αwe␈αdesign␈αan␈αinterpreter
␈↓ α←␈↓for␈αthe␈αlanguage␈αso␈αthat␈αthe␈αsystem␈αcan␈αuse␈αits␈αnew␈αknowledge␈αto␈α
keep␈αtrack
␈↓ α←␈↓of␈α⊂the␈α∂details␈α⊂of␈α⊂data␈α∂structure␈α⊂construction␈α⊂and␈α∂maintenance.␈α⊂ This␈α⊂is,␈α∂of
␈↓ α←␈↓course, easy to say but somewhat harder to do.  Some difficult questions arise,

␈↓"β␈↓ α←␈↓↓␈↓ β'What␈α
knowledge␈αabout␈α
its␈αrepresentations␈α
does␈αa␈α
system␈α
require␈αin
␈↓ α←␈↓↓␈↓ β'order␈α⊂to␈α⊂allow␈α⊂it␈α⊂to␈α⊂do␈α⊂a␈α⊂range␈α⊂of␈α⊂nontrivial␈α⊂management␈α∂tasks?
␈↓ α←␈↓↓␈↓ β'How␈α⊗should␈α⊗this␈α∃knowledge␈α⊗be␈α⊗organized?␈α∃ How␈α⊗should␈α⊗it␈α∃be
␈↓ α←␈↓↓␈↓ β'represented?  How can it be used?
␈↓ α←␈↓␈↓6-5␈↓ λQTHE SOLUTION    131␈↓

␈↓"β␈↓ α←␈↓All␈αthese␈αissues␈αare␈αdealt␈αwith␈αbelow.␈α We␈αdemonstrate,␈αfor␈αinstance,␈αthat␈αthe
␈↓ α←␈↓relevant␈α"knowledge␈α"includes␈α"information␈α"about␈α"the␈α"structure␈α"and
␈↓ α←␈↓interrelations␈αof␈α
representations␈αand␈αshow␈α
that␈αit␈α
can␈αbe␈αused␈α
as␈αthe␈αbasis␈α
for
␈↓ α←␈↓the interactive transfer of domain-specific expertise.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂main␈α∂task␈α∂here,␈α∂then,␈α∂is␈α∂the␈α∂description␈α∂and␈α∂use␈α∂of␈α∂knowledge
␈↓ α←␈↓about␈αrepresentations.␈α To␈αaccomplish␈αthis,␈αwe␈αuse␈αa␈α␈↓↓data␈αstructure␈αschema␈↓,␈αa
␈↓ α←␈↓device␈αthat␈α
provides␈αa␈αframework␈α
and␈αlanguage␈α
in␈αwhich␈αrepresentations␈α
can
␈↓ α←␈↓be␈α∂specified.␈α⊂ The␈α∂framework,␈α⊂like␈α∂most,␈α⊂carries␈α∂its␈α⊂own␈α∂perspective␈α⊂on␈α∂its
␈↓ α←␈↓domain.␈α⊂ One␈α⊂point␈α⊃it␈α⊂emphasizes␈α⊂strongly␈α⊃is␈α⊂the␈α⊂detailed␈α⊃specification␈α⊂of
␈↓ α←␈↓many␈α∞kinds␈α∞of␈α∞information␈α∞about␈α∞representations.␈α∞ It␈α∞attempts␈α∞to␈α∂make␈α∞this
␈↓ α←␈↓specification␈α⊂task␈α⊂easier␈α⊂by␈α∂providing␈α⊂an␈α⊂organization␈α⊂for␈α⊂the␈α∂information
␈↓ α←␈↓and a relatively high-level vocabulary for its expression.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈αthat␈αthe␈αschemata␈αform␈αthe␈αsecond␈αmajor␈αexample␈αof␈αmeta-level
␈↓ α←␈↓knowledge.␈α∞ While␈α∞a␈α∞particular␈α∞data␈α∞structure␈α∞may␈α∞be␈α∞used␈α∞to␈α∂represent␈α∞an
␈↓ α←␈↓object␈α∂in␈α∂the␈α∂domain,␈α∂the␈α∂schemata␈α∂(as␈α∂descriptions␈α∂of␈α∂representations)␈α∂are
␈↓ α←␈↓meta-level objects.

␈↓"β␈↓ α←␈↓␈↓α6-6    KEY IDEAS:  COMMENTS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?To␈α∪provide␈α∪some␈α∪background␈α∩for␈α∪understanding␈α∪the␈α∪examples␈α∩of
␈↓ α←␈↓system␈α∂performance␈α∂that␈α∂follow,␈α∂we␈α∞present␈α∂below␈α∂some␈α∂brief␈α∂comments␈α∞on
␈↓ α←␈↓several of the ideas listed in Section 6-2.

␈↓"β␈↓ α←␈↓␈↓α6-6-1    Vocabulary␈↓
␈↓"β␈↓ α←␈↓␈↓ β?In␈αthe␈αdiscussion␈αthat␈αfollows,␈αthe␈αterms␈α␈↓↓data␈αstructure␈↓,␈α␈↓↓extended␈αdata
␈↓ α←␈↓↓type␈↓,␈αand␈α␈↓↓representation␈↓␈αwill␈αbe␈αused␈αinterchangeably.␈α Equating␈αthe␈αfirst␈αtwo
␈↓ α←␈↓implies␈α∞extending␈α∞the␈α∞idea␈α
of␈α∞data␈α∞types␈α∞to␈α
cover␈α∞every␈α∞data␈α∞structure␈α∞in␈α
a
␈↓ α←␈↓system.␈α∞ The␈α∞utility␈α∞of␈α∞this␈α∞view␈α∞appears␈α∞to␈α∞be␈α∞widely␈α∞accepted␈α∞and,␈α∂in␈α∞the
␈↓ α←␈↓case␈α∞at␈α∞hand,␈α
will␈α∞influence␈α∞our␈α
approach␈α∞to␈α∞determining␈α∞what␈α
information
␈↓ α←␈↓about␈α∃data␈α∃structures␈α∃is␈α∃relevant␈α∃and␈α∃how␈α∃that␈α∃information␈α∃should␈α∀be
␈↓ α←␈↓organized.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
equivalence␈α
of␈α
the␈α
last␈α
two␈α
suggests␈α
our␈α
perspective␈α
on␈α
the␈α
design
␈↓ α←␈↓and␈α∞implementation␈α
of␈α∞knowledge␈α
representations.␈α∞ These␈α∞two␈α
tasks--design
␈↓ α←␈↓and␈αimplementation--are␈αtypically␈αdecoupled,␈αand,␈αindeed,␈αthe␈αdesirability␈αof
␈↓ α←␈↓transparency␈α∞of␈α∞implementation␈α∞has␈α∞been␈α∞stressed␈α∞from␈α∞many␈α∞quarters␈α
(e.g.,
␈↓ α←␈↓[Bachman75],␈α~[Balzer67],␈α≠[Liskov74]).␈α~ But␈α~what␈α≠might␈α~we␈α≠learn␈α~by
␈↓ α←␈↓considering␈α_them␈α_simultaneously?␈α_ That␈α→is,␈α_what␈α_can␈α_we␈α→learn␈α_about
␈↓ α←␈↓representation␈α↔design␈α_by␈α↔considering␈α_issues␈α↔that␈α↔arise␈α_at␈α↔the␈α_level␈α↔of
␈↓ α←␈↓implementation␈α∞and␈α∞technical␈α∞detail?␈α∂ Conversely,␈α∞what␈α∞can␈α∞we␈α∂learn␈α∞about
␈↓ α←␈↓the␈α∩organization␈α∩or␈α⊃design␈α∩of␈α∩data␈α∩types␈α⊃by␈α∩viewing␈α∩them␈α∩as␈α⊃knowledge
␈↓ α←␈↓representations?  We examine these questions below.

␈↓"β␈↓ α←␈↓␈↓α6-6-2    Schemata as knowledge representation descriptions␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈α⊗noted,␈α⊗the␈α⊗schemata␈α↔are␈α⊗the␈α⊗primary␈α⊗vehicle␈α↔for␈α⊗describing
␈↓ α←␈↓representations.␈α
 They␈α∞were␈α
developed␈α∞as␈α
a␈α
generalization␈α∞of␈α
the␈α∞concept␈α
of
␈↓ α←␈↓record␈α∞structures␈α∞and␈α
strongly␈α∞resemble␈α∞them␈α
in␈α∞both␈α∞organization␈α∞and␈α
use.
␈↓ α←␈↓␈↓132    KNOWLEDGE ACQUISITION II␈↓ 
#6-6␈↓

␈↓"β␈↓ α←␈↓Many␈αof␈αthe␈αoperations␈αwith␈αthe␈αschemata␈αcan␈αbe␈αseen␈αin␈αterms␈αof␈α
variations
␈↓ α←␈↓on␈αthe␈αtask␈αof␈α
creating␈αa␈αnew␈αinstance␈αof␈α
a␈αrecord-like␈αstructure.␈α We␈αwill␈α
see
␈↓ α←␈↓that␈αthese␈αoperations␈αproceed␈αin␈αa␈αmixed-initiative␈αmode:␈αThe␈αneed␈αto␈αadd␈αa
␈↓ α←␈↓new␈α∂data␈α⊂structure␈α∂is␈α∂made␈α⊂evident␈α∂by␈α∂an␈α⊂action␈α∂on␈α∂the␈α⊂part␈α∂of␈α⊂the␈α∂user;
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓␈α
then␈α
takes␈α∞over,␈α
retrieving␈α
the␈α
appropriate␈α∞schema␈α
and␈α
using␈α∞it␈α
to
␈↓ α←␈↓guide the rest of the interaction.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
schemata␈α
take␈αfrom␈α
records␈α
the␈αconcept␈α
of␈α
structure␈αdescription,
␈↓ α←␈↓the␈α∞separation␈α
of␈α∞representation␈α
from␈α∞implementation,␈α
and␈α∞the␈α
fundamental
␈↓ α←␈↓record␈αcreation␈αoperation.␈α Records␈αprovide␈αa␈αsimple␈αlanguage␈αfor␈αdescribing
␈↓ α←␈↓data␈αstructures,␈αand␈αthis␈αwas␈αused␈αas␈αthe␈αbasis␈αfor␈αthe␈αstructure␈αsyntax␈αin␈αthe
␈↓ α←␈↓schemata.␈α≠ Records␈α≠also␈α≠isolate␈α≠conceptual␈α≠structures␈α≠from␈α≠details␈α≠of
␈↓ α←␈↓implementation.␈α⊃ Thus,␈α∩code␈α⊃may␈α⊃uniformly␈α∩refer␈α⊃to␈α⊃field␈α∩F␈α⊃of␈α∩record␈α⊃R
␈↓ α←␈↓despite␈α∃changes␈α⊗in␈α∃the␈α∃way␈α⊗the␈α∃record␈α∃is␈α⊗actually␈α∃stored.␈α⊗ Finally,␈α∃the
␈↓ α←␈↓operation␈α
of␈α
creating␈α
a␈αnew␈α
instance␈α
of␈α
a␈αrecord␈α
was␈α
used␈α
as␈αthe␈α
fundamental
␈↓ α←␈↓paradigm␈α
for␈α
this␈α
part␈α
of␈αthe␈α
knowledge␈α
acquisition␈α
task.␈α
 At␈α
the␈αglobal␈α
level,
␈↓ α←␈↓much␈α∪that␈α∩happens␈α∪in␈α∩this␈α∪chapter␈α∩can␈α∪be␈α∩viewed␈α∪in␈α∩terms␈α∪of␈α∩creating
␈↓ α←␈↓instances from one or more kinds of records.

␈↓"β␈↓ α←␈↓␈↓αExtensions--data structure syntax␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈αbasic␈αidea␈αof␈αa␈α
record-like␈αdescriptor␈αwas␈αthen␈αextended␈αto␈α
make
␈↓ α←␈↓possible␈α∞the␈α∞capabilities␈α
we␈α∞require.␈α∞ The␈α
structure␈α∞syntax␈α∞was␈α∞extended␈α
by
␈↓ α←␈↓adopting␈αsome␈αof␈αthe␈αconventions␈αof␈αBNF,␈αso␈αthat␈αa␈αcertain␈αvariability␈αcould
␈↓ α←␈↓be␈α⊂described.␈α⊃ For␈α⊂instance,␈α⊃a␈α⊂schema␈α⊂can␈α⊃indicate␈α⊂that␈α⊃a␈α⊂structure␈α⊃has␈α⊂␈↓↓a
␈↓ α←␈↓↓minimum of 1, a maximum of 4, and typically 2␈↓ components of a given form.␈↓
2␈↓

␈↓"β␈↓ α←␈↓␈↓αExtensions--data structure interrelationships␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?In␈α∂addition,␈α∞we␈α∂introduced␈α∞a␈α∂syntax␈α∞of␈α∂data␈α∂structure␈α∞interrelations.
␈↓ α←␈↓As␈α
noted␈α
above,␈α
data␈α
structures␈α
in␈α
a␈α
program␈α
are␈α
not␈α
typically␈α
independent
␈↓ α←␈↓and␈α∂the␈α∂addition␈α∂of␈α∞a␈α∂new␈α∂instance␈α∂of␈α∂some␈α∞data␈α∂type␈α∂to␈α∂the␈α∂system␈α∞often
␈↓ α←␈↓requires extensive bookkeeping to maintain the existing interdependencies.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α≠problem␈α≠has␈α≠been␈α≠considered␈α≠previously,␈α≠primarily␈α≠with
␈↓ α←␈↓techniques␈α
oriented␈α∞around␈α
demon-like␈α
mechanisms␈α∞(e.g.,␈α
the␈α
demons␈α∞in␈α
␈↓¬QA4␈↓
␈↓ α←␈↓[Rulifson72]).␈α The␈αapproach␈αtaken␈αhere␈α
differs␈αin␈αseveral␈αrespects.␈αWhile,␈α
as
␈↓ α←␈↓in␈α∀previous␈α∀approaches,␈α∀demon-like␈α∀mechanisms␈α∀were␈α∀employed␈α∀to␈α∀help
␈↓ α←␈↓model␈α⊂the␈α⊂domain,␈α⊂they␈α∂will␈α⊂also␈α⊂be␈α⊂used␈α∂extensively␈α⊂at␈α⊂the␈α⊂level␈α⊂of␈α∂data
␈↓ α←␈↓structures,␈α⊂as␈α⊃a␈α⊂tool␈α⊃to␈α⊂aid␈α⊂in␈α⊃management␈α⊂of␈α⊃the␈α⊂knowledge␈α⊃base.␈α⊂ They
␈↓ α←␈↓become␈αan␈αimportant␈αcomponent␈αof␈αour␈αrepresentation␈αmethodology␈αand␈αwill
␈↓ α←␈↓be seen to have an influence on the organization of knowledge in the system.
␈↓"β␈↓ α←␈↓␈↓ β?Previous␈α∂uses␈α⊂of␈α∂demons␈α⊂have␈α∂also␈α⊂involved␈α∂the␈α⊂full␈α∂power␈α⊂of␈α∂the
␈↓ α←␈↓parent␈α⊃programming␈α⊃language,␈α⊃as␈α⊃in␈α⊃␈↓¬QA4␈↓␈α⊃or␈α⊃␈↓¬PLANNER␈↓,␈α⊃where␈α⊃the␈α⊃body␈α⊃of␈α⊂a
␈↓ α←␈↓demon␈α⊃can␈α⊃be␈α⊃an␈α⊃arbitrary␈α⊃computation.␈α⊃ For␈α⊃reasons␈α⊃which␈α⊃will␈α⊂become

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[2]␈α∃This␈α∃variability␈α∀is␈α∃what␈α∃led␈α∃to␈α∀calling␈α∃them␈α∃␈↓↓schemata␈↓,␈α∃rather␈α∀than
␈↓ α←␈↓declarations␈αor␈α
records,␈αsince␈α
the␈αlatter␈α
typically␈αdescribe␈α
structures␈αwith␈α
fixed
␈↓ α←␈↓formats.
␈↓ α←␈↓␈↓6-6␈↓ π\KEY IDEAS:  COMMENTS    133␈↓

␈↓"β␈↓ α←␈↓clear␈αlater,␈αsignificant␈α
effort␈αwas␈αput␈αinto␈α
avoiding␈αthis␈αapproach.␈α
 We␈αhave
␈↓ α←␈↓instead␈α∩developed␈α∩a␈α∩small␈α∪syntax␈α∩of␈α∩interrelationships␈α∩that␈α∪expresses␈α∩the
␈↓ α←␈↓relevant facts in a straightforward form.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α_fundamental␈α_point␈α_here␈α_is␈α_simple␈α_enough: ␈α_Whatever␈α↔the
␈↓ α←␈↓interrelationships,␈α≠they␈α~should␈α≠be␈α~made␈α≠explicit.␈α~ All␈α≠too␈α≠often,␈α~the
␈↓ α←␈↓interdependencies␈α
of␈αinternal␈α
data␈αstructures␈α
are␈αleft␈α
either␈αas␈α
folklore␈α
or,␈αat
␈↓ α←␈↓best,␈α∞mentioned␈α
briefly␈α∞in␈α∞documentation.␈α
In␈α∞line␈α∞with␈α
the␈α∞major␈α∞themes␈α
of
␈↓ α←␈↓this␈α⊂work,␈α⊂we␈α⊂want␈α⊂to␈α⊂make␈α⊂this␈α⊂knowledge␈α⊂explicit␈α⊂and␈α⊂accessible␈α⊃to␈α⊂the
␈↓ α←␈↓system itself.  The interrelationship syntax was the tool employed to do this.

␈↓"β␈↓ α←␈↓␈↓α6-6-3    A ``totally typed'' language␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
basic␈α∞approach,␈α
then,␈α∞was␈α
to␈α∞view␈α
the␈α∞representation␈α
primitives
␈↓ α←␈↓as␈α
extended␈α
data␈α
types␈α
in␈α
a␈α
high-level␈α
language,␈α
use␈α
an␈α
augmented␈α
record-
␈↓ α←␈↓like␈α∪structure␈α∩to␈α∪describe␈α∪each␈α∩of␈α∪them,␈α∪and␈α∩then␈α∪make␈α∪those␈α∩structures
␈↓ α←␈↓available␈αfor␈αreference␈αby␈αthe␈αsystem␈α
itself.␈α The␈αnext␈αstep␈αwas␈αto␈α
apply␈αthis
␈↓ α←␈↓exhaustively␈α∪and␈α∪uniformly␈α∩to␈α∪every␈α∪object␈α∪in␈α∩the␈α∪system.␈α∪ That␈α∪is,␈α∩the
␈↓ α←␈↓``language''␈α
should␈α
be␈α
``totally␈α
typed,''␈α
and␈α
every␈α
object␈α
in␈α
the␈α
system␈α
should␈α
be
␈↓ α←␈↓an␈α
instance␈α
of␈α
some␈α∞schema.␈α
 One␈α
reason␈α
for␈α∞this␈α
is␈α
data␈α
base␈α∞integrity.␈α
 A
␈↓ α←␈↓totally␈αtyped␈αlanguage␈αmakes␈αpossible␈α
exhaustive␈αtype␈αchecking␈αand␈αone␈α
level
␈↓ α←␈↓of␈α∞knowledge␈α∞base␈α∞integrity.␈α∞ In␈α∂addition,␈α∞since␈α∞many␈α∞of␈α∞the␈α∂extended␈α∞data
␈↓ α←␈↓types␈αcorrespond␈αto␈αdomain-specific␈α
objects,␈αthe␈αknowledge␈αacquisition␈α
dialog
␈↓ α←␈↓can␈αbe␈αmade␈αto␈αappear␈αto␈αthe␈αexpert␈αto␈αbe␈αphrased␈αin␈αterms␈αof␈αobjects␈αin␈α
the
␈↓ α←␈↓domain,␈α⊃while␈α⊂to␈α⊃the␈α⊂system␈α⊃it␈α⊂is␈α⊃a␈α⊂straightforward␈α⊃manipulation␈α⊃of␈α⊂data
␈↓ α←␈↓structures.␈α∂ It␈α⊂thus␈α∂helps␈α⊂bridge␈α∂the␈α⊂gap␈α∂in␈α⊂perspectives.␈α∂ Finally,␈α⊂since␈α∂we
␈↓ α←␈↓were␈α∂concerned␈α∂with␈α∂the␈α⊂large␈α∂amount␈α∂of␈α∂knowledge␈α⊂about␈α∂representations
␈↓ α←␈↓that␈αis␈αtypically␈αleft␈αimplicit,␈αapplying␈αthe␈αschema␈αidea␈αto␈αevery␈αobject␈αin␈αthe
␈↓ α←␈↓system␈α∩offered␈α∪some␈α∩level␈α∩of␈α∪assurance␈α∩that␈α∩we␈α∪had␈α∩made␈α∪explicit␈α∩some
␈↓ α←␈↓significant fraction of this information.
␈↓"β␈↓ α←␈↓␈↓ β?Exhaustive␈α∨application␈α∨of␈α∨the␈α∨schema␈α∨idea␈α presents␈α∨several
␈↓ α←␈↓implications.␈α First,␈αit␈αmeans␈αthat␈αeven␈αthe␈αcomponents␈αfrom␈αwhich␈αa␈αschema
␈↓ α←␈↓is␈αbuilt␈αshould␈αalso␈αbe␈αinstances␈αof␈αsome␈α(other)␈αschema,␈αand␈αwe␈αwill␈αsee␈αthat
␈↓ α←␈↓this␈αis␈αtrue.␈α Second,␈αsince␈αwe␈αclaim␈αboth␈αthat␈αthe␈αschemata␈αshould␈αbe␈αa␈αpart
␈↓ α←␈↓of␈α∞the␈α
system␈α∞and␈α
that␈α∞every␈α
object␈α∞in␈α
the␈α∞system␈α
should␈α∞be␈α
an␈α∞instance␈α
of
␈↓ α←␈↓some␈α⊗schema,␈α⊗then␈α⊗the␈α⊗schemata␈α∃themselves␈α⊗should␈α⊗be␈α⊗an␈α⊗instance␈α∃of
␈↓ α←␈↓something.␈α⊃ In␈α⊃more␈α⊂familiar␈α⊃terms,␈α⊃␈↓↓if␈α⊃the␈α⊂structure␈α⊃declarations␈α⊃are␈α⊃to␈α⊂be
␈↓ α←␈↓↓objects␈αin␈αthe␈αprogram,␈αand␈αif␈αeverything␈αis␈α
to␈αbe␈αa␈αdata␈αtype␈αof␈αsome␈αsort,␈α
then
␈↓ α←␈↓↓the␈α
declarations␈α∞themselves␈α
must␈α
be␈α∞a␈α
data␈α∞type␈↓.␈α
This␈α
was␈α∞done.␈α
 There␈α∞is␈α
a
␈↓ α←␈↓``schema-schema,''␈αwhich␈αspecifies␈α
the␈αstructure␈αof␈α
a␈αschema,␈αand␈αall␈α
schemata
␈↓ α←␈↓are instances of it.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈α∞the␈α∞schema-schema␈α∞indicates␈α∞the␈α
structure␈α∞of␈α∞a␈α∞schema,␈α∞it␈α
can
␈↓ α←␈↓be␈α∩used␈α∩to␈α∩guide␈α∩the␈α∩creation␈α∩of␈α∩new␈α∩data␈α∩types.␈α∩ This␈α∩offers␈α∩the␈α⊃same
␈↓ α←␈↓benefits␈αas␈αbefore,␈αof␈αa␈αcertain␈αlevel␈αof␈αintegrity␈αand␈αa␈αrelatively␈α``high-level''
␈↓ α←␈↓dialog.␈α⊃ Note␈α⊃that␈α∩it␈α⊃deals,␈α⊃however,␈α⊃with␈α∩the␈α⊃fairly␈α⊃sophisticated␈α∩task␈α⊃of
␈↓ α←␈↓specifying a new data structure.
␈↓"β␈↓ α←␈↓␈↓ β?While␈α⊂the␈α⊂recursive␈α⊂application␈α⊃of␈α⊂the␈α⊂schema␈α⊂idea␈α⊃was␈α⊂motivated
␈↓ α←␈↓␈↓134    KNOWLEDGE ACQUISITION II␈↓ 
#6-6␈↓

␈↓"β␈↓ α←␈↓initially␈α
by␈α
purely␈α∞utilitarian␈α
considerations,␈α
it␈α
led␈α∞to␈α
a␈α
useful␈α∞uniformity␈α
in
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓: ␈α
There␈αis␈α
a␈α
single␈αprocess␈α
by␈α
which␈αboth␈α
the␈α
schema-schema␈αcan␈α
be
␈↓ α←␈↓instantiated␈αto␈αcreate␈αa␈αnew␈αschema␈α(a␈αnew␈αknowledge␈αrepresentation)␈αand␈α
by
␈↓ α←␈↓which␈α
a␈α
schema␈α∞can␈α
be␈α
instantiated␈α∞to␈α
create␈α
a␈α∞new␈α
instance␈α
of␈α∞an␈α
existing
␈↓ α←␈↓knowledge␈α⊂representation.␈α⊃ This␈α⊂not␈α⊂only␈α⊃made␈α⊂possible␈α⊃bootstrapping␈α⊂the
␈↓ α←␈↓system␈α∀(described␈α∀later␈α∃in␈α∀this␈α∀chapter)␈α∀but␈α∃also␈α∀supplied␈α∀much␈α∃of␈α∀the
␈↓ α←␈↓generality␈α
of␈α
the␈α∞approach.␈α
 Part␈α
of␈α
Fig.␈α∞4-1␈α
has␈α
been␈α
reproduced␈α∞below␈α
to
␈↓ α←␈↓illustrate this multi-level organization.

␈↓"␈↓ α←␈↓∧                      KNOWLEDGE ACQUISITION
␈↓"␈↓ α←␈↓∧            ⊂ααααααααααααααααααααααααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            ~ ⊂παααααααααααααααπ⊃                   ~
␈↓"␈↓ α←␈↓∧            ~ ~~2              ~~α α α α α ⊃        ~
␈↓"␈↓ α←␈↓∧            ~ ~~ schema-schema ~~          ↓        ~
␈↓"␈↓ α←␈↓∧            ~ %∀ααααααααααααααα∀$ ⊂αααααααααααααα⊃  ~
␈↓"␈↓ α←␈↓∧            ~         ⊂ α α α α α ~ new schema   ~  ~
␈↓"␈↓ α←␈↓∧            ~         ↓           ~ acquisition  ~  ~
␈↓"␈↓ α←␈↓∧            ~ ⊂παααααααααααααααπ⊃ %αααααααααααααα$  ~
␈↓"␈↓ α←␈↓∧            ~ ~~1              ~~α α α α α ⊃        ~
␈↓"␈↓ α←␈↓∧            ~ ~~    schemata   ~~          ↓        ~
␈↓"␈↓ α←␈↓∧ KNOWLEDGE  ~ %∀ααααααααααααααα∀$ ⊂αααααααααααααα⊃  ~
␈↓"␈↓ α←␈↓∧   BASE     ~                     ~ new instance ~  ~
␈↓"␈↓ α←␈↓∧⊂πααααααπ⊃  ~                ⊂ α α~ acquisition  ~  ~← α EXPERT
␈↓"␈↓ α←␈↓∧~~0     ~~←α~α α α α α α α α $    %αααααααααααααα$  ~[dialog]
␈↓"␈↓ α←␈↓∧~~ facts~~  ~    [knowledge                         ~
␈↓"␈↓ α←␈↓∧~~ -----~~  ~     transfer]                         ~
␈↓"␈↓ α←␈↓∧~~ rules~~  ~                                       ~
␈↓"␈↓ α←␈↓∧~~      ~~  ~                                       ~
␈↓"␈↓ α←␈↓∧%∀αααααα∀$  ~                                       ~
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            ~                                       ~
␈↓"␈↓ α←␈↓∧            %ααααααααααααααααααααααααααααααααααααααα$


␈↓"␈↓ α←␈↓α␈↓ β→Fig. 6-1.    The multi-level process of acquiring new primitives.    


␈↓"β␈↓ α←␈↓␈↓α6-6-4    Knowledge base integrity␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Avoiding␈α∞bugs␈α∞when␈α∞manipulating␈α∂data␈α∞structures␈α∞is␈α∞also␈α∂known␈α∞as
␈↓ α←␈↓``assuring␈α
the␈αintegrity␈α
of␈α
the␈αdata␈α
base''␈α
and␈αhas␈α
been␈α
investigated␈αwithin␈α
the
␈↓ α←␈↓framework␈α⊃of␈α⊃several␈α∩organizational␈α⊃paradigms␈α⊃(see␈α⊃e.g.,␈α∩[McLeod76]␈α⊃and
␈↓ α←␈↓[Eswarn75]).␈α
 Previous␈α
efforts␈α
have␈αemphasized␈α
the␈α
utility␈α
of␈α
extensive␈αtype
␈↓ α←␈↓checking␈α∀for␈α∪extended␈α∀data␈α∀types␈α∪and␈α∀have␈α∪studied␈α∀aspects␈α∀of␈α∪integrity
␈↓ α←␈↓specific␈α∂to␈α∂a␈α∂particular␈α∂paradigm.␈α∂ We␈α∂use␈α∂many␈α∂of␈α∂these␈α∂same␈α∂techniques
␈↓ α←␈↓here,␈α
but␈α
focus␈αon␈α
the␈α
problem␈αof␈α
interrelationships␈α
between␈α
data␈αstructures
␈↓ α←␈↓in␈α∂general␈α∂and␈α∞concentrate␈α∂on␈α∂dealing␈α∞with␈α∂the␈α∂effects␈α∞of␈α∂additions␈α∂on␈α∞the
␈↓ α←␈↓integrity of the knowledge base.
␈↓ α←␈↓␈↓6-6␈↓ π\KEY IDEAS:  COMMENTS    135␈↓

␈↓"β␈↓ α←␈↓␈↓ β?While␈α∂it␈α∞has␈α∂not␈α∞been␈α∂possible␈α∂to␈α∞devise␈α∂ways␈α∞of␈α∂assuring␈α∂the␈α∞total
␈↓ α←␈↓integrity␈αof␈αthe␈αknowledge␈αbase,␈αthe␈α
capabilities␈αof␈αour␈αsystem␈αcan␈αbe␈α
broadly
␈↓ α←␈↓classified␈α∞by␈α∞considering␈α∞three␈α∞error␈α∞sources.␈α∞ First,␈α∞the␈α∞system␈α∞can␈α∞assure␈α
a
␈↓ α←␈↓form␈α∂of␈α∞completeness,␈α∂by␈α∞making␈α∂sure␈α∞both␈α∂that␈α∞the␈α∂expert␈α∞is␈α∂reminded␈α∞to
␈↓ α←␈↓supply␈α→every␈α→necessary␈α_component␈α→of␈α→a␈α_structure␈α→and␈α→that␈α→all␈α_other
␈↓ α←␈↓appropriate␈α
structures␈α∞are␈α
informed␈α∞of␈α
the␈α
newly␈α∞added␈α
item␈α∞(``informed''␈α
is
␈↓ α←␈↓elaborated␈α∃below).␈α∀ Second,␈α∃it␈α∀can␈α∃assure␈α∀``syntactic''␈α∃integrity.␈α∃ There␈α∀is
␈↓ α←␈↓complete␈α⊃type␈α⊃checking,␈α⊃and␈α⊃no␈α⊃interaction␈α⊃with␈α⊃the␈α⊃expert␈α⊃will␈α∩result␈α⊃in
␈↓ α←␈↓incorrect data types in the knowledge base.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α∩it␈α∩can␈α∩assure␈α∩a␈α∩certain␈α∩level␈α∩of␈α∩``semantic''␈α∩integrity.␈α∩ The
␈↓ α←␈↓semantics␈α∂of␈α∂any␈α∂individual␈α∂data␈α∂structure␈α∂will␈α∂be␈α∂properly␈α⊂maintained,␈α∂so
␈↓ α←␈↓that,␈α
for␈α
instance,␈α
a␈α
new␈α
attribute␈α
will␈α
be␈α
given␈α
all␈α
the␈αdescriptors␈α
appropriate
␈↓ α←␈↓to␈αit,␈αin␈αthe␈αcorrect␈αform.␈α It␈αcan␈αalso␈αassure␈αsome␈αsemantic␈αconsistency␈αin␈αtwo
␈↓ α←␈↓or␈α_more␈α_related␈α↔structures,␈α_but␈α_this␈α_is␈α↔as␈α_yet␈α_incomplete,␈α_since␈α↔some
␈↓ α←␈↓inconsistencies␈αcan␈αarise␈α
that␈αrequire␈αmore␈α
knowledge␈αabout␈αthe␈αdomain␈α
than
␈↓ α←␈↓is␈αcurrently␈αavailable.␈α For␈αinstance,␈αin␈αthe␈αmedical␈αdomain,␈αwhile␈αdescribing
␈↓ α←␈↓a␈αnew␈αorganism␈αa␈α
user␈αmight␈αindicate␈αthat␈α
it␈αis␈αan␈α``acid-fast␈α
coccus''␈α(``acid-
␈↓ α←␈↓fast''␈α
describes␈α
a␈αresponse␈α
to␈α
a␈αkind␈α
of␈α
stain)␈α
when,␈αin␈α
fact,␈α
the␈αcombination␈α
is
␈↓ α←␈↓biologically␈α≠meaningless.␈α≠ Each␈α≠individual␈α≠answer␈α≠is␈α≠correct␈α≤but␈α≠the
␈↓ α←␈↓combination is inconsistent for reasons that are not easily represented.

␈↓"β␈↓ α←␈↓␈↓α6-6-5    Summary␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αschemata␈αand␈αtheir␈αassociated␈αstructures␈αprovide␈αa␈αlanguage␈αand
␈↓ α←␈↓framework␈α~in␈α~which␈α~representations␈α~can␈α~be␈α~specified.␈α~ It␈α~should␈α~be
␈↓ α←␈↓emphasized␈αthat␈αall␈αof␈αthe␈αwork␈αreported␈αhere␈αwas␈αat␈αthe␈αlevel␈αof␈αthe␈αdesign
␈↓ α←␈↓and␈α
the␈αimplementation␈α
of␈αthis␈α
language␈αand␈α
framework␈αin␈α
␈↓¬TEIRESIAS␈↓.␈α Some
␈↓ α←␈↓of␈α∞the␈α∞representations␈α∞that␈α
the␈α∞language␈α∞can␈α∞describe␈α
are␈α∞those␈α∞used␈α∞in␈α
the
␈↓ α←␈↓current␈α∞performance␈α∂program␈α∞(␈↓¬MYCIN␈↓);␈α∂later␈α∞sections␈α∂of␈α∞this␈α∂chapter␈α∞examine
␈↓ α←␈↓the␈αlimits␈αof␈αits␈αexpressive␈αpower.␈α Within␈αthose␈αlimits,␈αthe␈αsystem␈αdeals␈αwith
␈↓ α←␈↓the␈αgeneral␈αissue␈αof␈αthe␈αdesign␈αand␈αspecification␈αof␈αrepresentations.␈α Nothing
␈↓ α←␈↓here␈α∂is␈α∞specific␈α∂to␈α∞medicine␈α∂or␈α∞to␈α∂the␈α∂attribute-object-value␈α∞representations
␈↓ α←␈↓that␈α⊃we␈α⊃will␈α⊂see␈α⊃employed.␈α⊃ Within␈α⊃the␈α⊂range␈α⊃of␈α⊃representations␈α⊃that␈α⊂our
␈↓ α←␈↓framework␈α∞permits,␈α∞the␈α∞system␈α∞is␈α∞domain␈α∞independent␈α∞and␈α∞has␈α∞a␈α∞degree␈α
of
␈↓ α←␈↓representation␈α_independence␈α_as␈α_well.␈α↔ This␈α_generality␈α_results␈α_from␈α↔the
␈↓ α←␈↓isolation␈α∞and␈α
stratification␈α∞of␈α
the␈α∞three␈α
different␈α∞levels␈α
of␈α∞knowledge␈α∞in␈α
the
␈↓ α←␈↓system, discussed in detail in Section 6-13.
␈↓ α←␈↓␈↓136    KNOWLEDGE ACQUISITION II␈↓ 
#6-6␈↓

␈↓"β␈↓ α←␈↓␈↓α6-7    TRACE OF SYSTEM PERFORMANCE:  ACQUIRING NEW
␈↓ α←␈↓α␈↓ β3VALUES␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Two␈α∂examples--the␈α∞acquisition␈α∂of␈α∞new␈α∂values␈α∞for␈α∂organism␈α∞identity
␈↓ α←␈↓and␈α∪for␈α∩culture␈α∪site--will␈α∩provide␈α∪an␈α∩overview␈α∪of␈α∪␈↓¬TEIRESIAS␈↓'s␈α∩capabilities.
␈↓ α←␈↓This␈α
demonstration␈αuses␈α
a␈αversion␈α
of␈αthe␈α
performance␈αprogram␈α
with␈α
a␈αvery
␈↓ α←␈↓simple␈αknowledge␈α
base,␈αas␈α
it␈αmight␈αappear␈α
in␈αan␈α
early␈αstage␈α
of␈αdevelopment
␈↓ α←␈↓when it contains only a few attributes and a few values for each.
␈↓"β␈↓ α←␈↓␈↓ β?Some preliminary comments should be made about these examples.
␈↓"β␈↓ α←␈↓␈↓ β?First,␈α
since␈αwe␈α
will␈α
be␈αdealing␈α
with␈α
some␈αcomplex␈α
data␈αstructures␈α
from
␈↓ α←␈↓a␈αspecific␈αperformance␈α
program,␈αmuch␈αof␈αwhat␈α
happens␈αin␈αthe␈α
trace␈αderives
␈↓ α←␈↓from␈α∀implementation␈α∀conventions␈α∀that␈α∀are␈α∀part␈α∀of␈α∀that␈α∀program.␈α∀ Since
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓'s␈αacquisition␈αprocess␈αhas␈αto␈αbe␈αthorough,␈αit␈αtakes␈αcare␈αof␈αall␈αof␈α
them.
␈↓ α←␈↓The␈α⊃important␈α∩point␈α⊃to␈α⊃note␈α∩is␈α⊃not␈α⊃what␈α∩these␈α⊃conventions␈α⊃are␈α∩but␈α⊃that
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓ can deal with them.
␈↓"β␈↓ α←␈↓␈↓ β?Second,␈α∂the␈α∂dialogs␈α∂are␈α∂at␈α∂times␈α∂deceptively␈α∂simple.␈α∂This␈α∂is␈α∂in␈α∂part
␈↓ α←␈↓some␈α∂measure␈α∂of␈α∂success,␈α∂since␈α⊂we␈α∂have␈α∂managed␈α∂to␈α∂delegate␈α∂much␈α⊂of␈α∂the
␈↓ α←␈↓detail␈α
to␈α␈↓¬TEIRESIAS␈↓,␈α
which␈α
takes␈αcare␈α
of␈αit␈α
quietly␈α
in␈αthe␈α
background.␈α
 To␈αsee
␈↓ α←␈↓this␈α⊂point␈α⊃most␈α⊂clearly,␈α⊂consider␈α⊃after␈α⊂reviewing␈α⊂the␈α⊃traces␈α⊂the␈α⊃amount␈α⊂of
␈↓ α←␈↓work␈α⊂that␈α⊂would␈α⊂be␈α⊂needed␈α⊂to␈α⊃do␈α⊂the␈α⊂same␈α⊂tasks␈α⊂by␈α⊂hand: ␈α⊂There␈α⊃are␈α⊂a
␈↓ α←␈↓number␈α∞of␈α∞details␈α
of␈α∞system␈α∞construction␈α∞that␈α
would␈α∞have␈α∞to␈α∞be␈α
memorized
␈↓ α←␈↓and␈αa␈αsignificant␈αamount␈αof␈αeffort␈αexpended␈αto␈αcreate␈αand␈αedit␈αthe␈αstructures
␈↓ α←␈↓by␈α
hand.␈α Much␈α
more␈αwork␈α
would␈αbe␈α
involved␈αif␈α
it␈αwere␈α
necessary␈α
to␈αteach
␈↓ α←␈↓the␈α∞expert␈α∞(who␈α∞may␈α∞never␈α∞have␈α∞programmed␈α∞before)␈α∞how␈α∞to␈α∞do␈α∂the␈α∞same
␈↓ α←␈↓things.␈α ␈↓¬TEIRESIAS␈↓␈αattends␈αto␈αthe␈αdetails,␈αdoes␈αmost␈αof␈αthe␈αwork,␈αand␈αmaintains
␈↓ α←␈↓a relatively high-level dialog.

␈↓"β␈↓ α←␈↓␈↓α6-7-1    Acquisition of a new organism identity␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α→first␈α→example␈α_shows␈α→how␈α→␈↓¬TEIRESIAS␈↓␈α_guides␈α→the␈α→process␈α_of
␈↓ α←␈↓describing␈α
a␈α∞new␈α
kind␈α∞of␈α
organism.␈α∞It␈α
demonstrates␈α∞primarily␈α
the␈α∞ability␈α
of
␈↓ α←␈↓the␈αschemata␈αto␈αkeep␈αtrack␈αof␈αall␈αthe␈αstructural␈αdetails␈αin␈αa␈αdata␈αstructure.␈α It
␈↓ α←␈↓also illustrates some of the human engineering features of the system.

␈↓ α←␈↓	***************************************************************

␈↓ α←␈↓↓␈↓ βW{The␈α∂expert␈α∂is␈α∂in␈α∂the␈α∞middle␈α∂of␈α∂a␈α∂normal␈α∂rule␈α∂acquisition␈α∞sequence
␈↓ α←␈↓↓␈↓ βWwhen␈α∪he␈α∪mentions␈α∪an␈α∪organism␈α∪name␈α∪that␈α∪the␈α∪system␈α∀has␈α∪never
␈↓ α←␈↓↓␈↓ βWencountered before.}

␈↓ α←␈↓	The new rule will be called RULE384
␈↓ α←␈↓	  If 1 -␈↓α THE SITE OF THE CULTURE IS BLOOD␈↓	
␈↓ α←␈↓	     2 -␈↓α THE PORTAL OF ENTRY OF THE ORGANISM IS THE␈↓	
␈↓ α←␈↓	        ␈↓αURINARY TRACT␈↓	
␈↓ α←␈↓	     3 -␈↓α THE PATIENT HAS NOT HAD A GENITO-URINARY␈↓	
␈↓ α←␈↓	        ␈↓α MANIPULATIVE PROCEDURE␈↓	
␈↓ α←␈↓	     4 -␈↓α CYSTITIS IS NOT A PROBLEM FOR WHICH THE␈↓	
␈↓ α←␈↓	        ␈↓α PATIENT HAS BEEN TREATED␈↓	
␈↓ α←␈↓	     5 -
␈↓ α←␈↓	 Then
␈↓ α←␈↓␈↓6-7␈↓ πNACQUIRING NEW VALUES    137␈↓

␈↓"β␈↓ α←␈↓	     1 -␈↓α THE ORGANISM IS LIKELY (.4) TO BE AN E.COLI␈↓	

␈↓ α←␈↓	Sorry, but I don't understand
␈↓ α←␈↓	   THE ORGANISM IS LIKELY (.4) TO BE AN E.COLI
␈↓ α←␈↓	because these words are unknown:  E.COLI

␈↓ α←␈↓↓␈↓ βW{␈↓¬TEIRESIAS␈↓↓␈αfinds␈α
it␈αcannot␈α
generate␈αany␈αinterpretations␈α
of␈αthe␈α
text,␈αso
␈↓ α←␈↓↓␈↓ βWit␈α
reprints␈αthe␈α
original␈α
line␈αand␈α
specifies␈α
where␈αthe␈α
trouble␈α
is.␈α This
␈↓ α←␈↓↓␈↓ βWmakes␈αit␈αpossible␈α
for␈αthe␈αexpert␈α
to␈αfigure␈αout␈α
what␈αit␈αis␈α
that␈αis␈αnew␈α
to
␈↓ α←␈↓↓␈↓ βWthe system.}

␈↓ α←␈↓	Would you care to try rephrasing that?
␈↓ α←␈↓	++** ␈↓αY␈↓	
␈↓ α←␈↓	(Please use *'s to mark what you think might be a new attribute
␈↓ α←␈↓	 or a new value of a attribute.)

␈↓ α←␈↓	        1 -␈↓α  THE ORGANISM IS LIKELY (.4) TO BE AN * E.COLI *␈↓	

␈↓ α←␈↓↓␈↓ βW{While␈α∞the␈α
language␈α∞in␈α∞this␈α
example␈α∞is␈α∞very␈α
simple,␈α∞there␈α∞are␈α
more
␈↓ α←␈↓↓␈↓ βWcomplex␈αcases␈αwhere␈αit␈αwould␈αbe␈αquite␈αhard␈αto␈αdecide␈αfrom␈αthe␈αcontext
␈↓ α←␈↓↓␈↓ βWwhat␈αsort␈αof␈αnew␈α
item␈αthe␈αexpert␈αhad␈α
in␈αmind.␈α To␈αavoid␈αthe␈α
necessity
␈↓ α←␈↓↓␈↓ βWfor␈α⊃more␈α∩sophisticated␈α⊃natural␈α∩language␈α⊃handling,␈α∩he␈α⊃is␈α∩asked␈α⊃to
␈↓ α←␈↓↓␈↓ βWmark off the new word or phrase with asterisks ... }

␈↓ α←␈↓	It looks as though
␈↓ α←␈↓	     E.COLI
␈↓ α←␈↓	refers to a new value of an attribute, correct?
␈↓ α←␈↓	++** ␈↓αY␈↓	
␈↓ α←␈↓	 ...is it a new value of the identity of a organism?
␈↓ α←␈↓	++** ␈↓αY␈↓	

␈↓ α←␈↓↓␈↓ βW{and␈α
with␈αthe␈α
new␈αitem␈α
clearly␈αmarked,␈α
the␈αsystem␈α
can␈αthen␈α
do␈αsome
␈↓ α←␈↓↓␈↓ βWvery simple pattern matching to decide what's going on.}

␈↓ α←␈↓	  Now please tell me a few things about it...

␈↓ α←␈↓↓␈↓ βW{␈↓¬TEIRESIAS␈↓↓␈α
retrieves␈α
the␈α
schema␈αfor␈α
an␈α
organism␈α
identity␈α
and␈α``reads''
␈↓ α←␈↓↓␈↓ βWit to find out what it needs to know about any organism.}

␈↓ α←␈↓	  Please give the full, formal name for "E.COLI"
␈↓ α←␈↓	  ++** ␈↓αESCHERICEA COLI␈↓	
␈↓ α←␈↓	  = ESCHERICEA-COLI

␈↓ α←␈↓↓␈↓ βW{Many␈α
long␈α
medical␈α
terms␈α
have␈α
abbreviations,␈α
so␈α
the␈α∞system␈α
requests
␈↓ α←␈↓↓␈↓ βWthe␈α∪formal␈α∪name␈α∪first.␈α∪ (Hyphens␈α∪are␈α∪inserted␈α∪to␈α∪make␈α∪a␈α∩single
␈↓ α←␈↓↓␈↓ βWidentifier out of it.)}

␈↓ α←␈↓	  Now please give all synonyms or abbreviations for ESCHERICEA-
␈↓ α←␈↓␈↓138    KNOWLEDGE ACQUISITION II␈↓ 
#6-7␈↓

␈↓"β␈↓ α←␈↓	  COLI which you would like the system to accept:
␈↓ α←␈↓	  [type an empty line when done]
␈↓ α←␈↓	  ++** ␈↓αE.COLI␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{The␈α
system␈αrequests␈α
synonyms␈α
and␈αabbreviations␈α
so␈αthat␈α
they␈α
can␈αbe
␈↓ α←␈↓↓␈↓ βWused␈α
in␈α
the␈α
future.␈α The␈α
first␈α
of␈α
them␈α
is␈αthen␈α
used␈α
through␈α
the␈αrest␈α
of
␈↓ α←␈↓↓␈↓ βWthe dialogue.}

␈↓ α←␈↓	  Please give a short description of E.COLI as a identity of a
␈↓ α←␈↓	  organism.
␈↓ α←␈↓	  ++** ␈↓α?␈↓	
␈↓ α←␈↓	  [e.g., for PSEUDOMONAS, the description is "AN ORGANISM
␈↓ α←␈↓	   COMMONLY FOUND IN THE HOSPITAL, QUITE OFTEN A SOURCE OF
␈↓ α←␈↓	   SECONDARY INFECTIONS IN PATIENTS WHOSE IMMUNE RESPONSE IS
␈↓ α←␈↓	   WEAKENED"]
␈↓ α←␈↓	  [type an empty line when done]
␈↓ α←␈↓	  ++** ␈↓αNORMAL FLORA IN THE LOWER INTESTINE, BUT A␈↓	
␈↓ α←␈↓	  ++** ␈↓αCAUSE OF SEPTICEMIA WHEN TRANSMITTED TO THE␈↓	
␈↓ α←␈↓	  ++** ␈↓αBLOOD STREAM␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{Two␈α∪useful␈α∩capabilities␈α∪are␈α∩displayed␈α∪here.␈α∩ First,␈α∪the␈α∪expert␈α∩is
␈↓ α←␈↓↓␈↓ βWasked␈α
to␈α
describe␈α
every␈α
structure␈α
he␈α
creates,␈α
providing␈α
a␈α
form␈α
of␈αon-
␈↓ α←␈↓↓␈↓ βWline␈α∞documentation␈α∞for␈α∞all␈α
data␈α∞structures.␈α∞ Since␈α∞the␈α
documentation
␈↓ α←␈↓↓␈↓ βWis␈α∞acquired␈α∞incrementally␈α
along␈α∞with␈α∞the␈α
rest␈α∞of␈α∞the␈α∞knowledge␈α
base,
␈↓ α←␈↓↓␈↓ βWthe␈αtask␈α
is␈αmuch␈αeasier␈α
than␈αattempting␈αto␈α
document␈αan␈αentire␈α
system
␈↓ α←␈↓↓␈↓ βWat once.
␈↓ α←␈↓↓␈↓ βWSecond,␈α_␈↓¬TEIRESIAS␈↓↓␈α_can␈α_generate␈α_sample␈α_answers␈α_for␈α_any␈α_of␈α↔its
␈↓ α←␈↓↓␈↓ βWquestions.␈α
 In␈αsome␈α
previous␈αsession,␈α
the␈αexpert␈α
had␈α
evidently␈αtaught
␈↓ α←␈↓↓␈↓ βWthe␈α
system␈αabout␈α
pseudomonas.␈α Since␈α
each␈αschema␈α
has␈α
an␈αindication
␈↓ α←␈↓↓␈↓ βWof␈αall␈αof␈αthe␈αinstances␈αthat␈αwere␈αcreated␈αfrom␈αit,␈αthe␈αsystem␈αpicks␈αone
␈↓ α←␈↓↓␈↓ βWof␈α∞those␈α∞instances␈α
and␈α∞displays␈α∞the␈α∞answer␈α
to␈α∞this␈α∞question␈α∞for␈α
that
␈↓ α←␈↓↓␈↓ βWinstance.}

␈↓ α←␈↓	  What is the aerobicity of a E.COLI?
␈↓ α←␈↓	  The answer should be in the form of 1 or more of the
␈↓ α←␈↓	  following-
␈↓ α←␈↓	           <aerobicity> <certainty factor>
␈↓ α←␈↓	  [Type 1 set to a line, then an empty line when done.]
␈↓ α←␈↓	  ++** ␈↓αFACULTATIVE .8␈↓	
␈↓ α←␈↓	  ++** ␈↓αAEROBIC .2␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{Each␈αof␈αthe␈αquestions␈αfrom␈αthe␈αsystem␈αis␈αprompted␈αby␈αthe␈αattempt␈αto
␈↓ α←␈↓↓␈↓ βWfill␈αin␈αsome␈αcomponent␈αof␈αthe␈αdata␈αstructure␈αbeing␈αbuilt,␈α
according␈αto
␈↓ α←␈↓↓␈↓ βWthe␈αrepresentation␈αconventions␈αindicated␈α
by␈αthe␈αschema.␈α One␈α
of␈αthose
␈↓ α←␈↓↓␈↓ βWconventions␈α∞indicates␈α∞that␈α∞aerobicity,␈α∞gramstain,␈α∞and␈α∂morphology␈α∞are
␈↓ α←␈↓␈↓6-7␈↓ πNACQUIRING NEW VALUES    139␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βWstored␈αas␈αpart␈αof␈αan␈αorganism.␈α The␈αschema␈αalso␈αindicates␈αthe␈αformat
␈↓ α←␈↓↓␈↓ βWfor␈αeach␈αpiece␈αof␈αsubstructure,␈α
and␈αthis␈αis␈αdisplayed␈αas␈αinstructions␈α
to
␈↓ α←␈↓↓␈↓ βWthe user.}

␈↓ α←␈↓	  What is the gramstain of a E.COLI?
␈↓ α←␈↓	  The answer should be in the form of a
␈↓ α←␈↓	           <gramstain>
␈↓ α←␈↓	  [Type 1 set to a line, then an empty line when done.]
␈↓ α←␈↓	  ++** ␈↓αROD␈↓	
␈↓ α←␈↓	    Sorry, but the following are invalid -
␈↓ α←␈↓	        ROD is not a recognized <gramstain>
␈↓ α←␈↓	    Please answer again [use the same answer if you really
␈↓ α←␈↓	    meant it.]
␈↓ α←␈↓	  ++** ␈↓αGRAMNEG␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{The␈α∂formatting␈α∂information␈α⊂also␈α∂allows␈α∂a␈α⊂check␈α∂on␈α∂the␈α⊂validity␈α∂of
␈↓ α←␈↓↓␈↓ βWeach␈α∞answer,␈α∞to␈α∂insure␈α∞that␈α∞all␈α∂information␈α∞added␈α∞to␈α∂the␈α∞knowledge
␈↓ α←␈↓↓␈↓ βWbase␈αis␈αproperly␈αstructured.␈α This␈αtime␈αthe␈αexpert␈αmade␈αa␈α
mistake.␈αIt
␈↓ α←␈↓↓␈↓ βWcan␈α
happen,␈α
however,␈α
that␈α
the␈α
answer␈α
is␈α
correct␈α
but␈α
the␈αperformance
␈↓ α←␈↓↓␈↓ βWprogram␈α
simply␈αhasn't␈α
heard␈αof␈α
it␈αyet.␈α
 Early␈αin␈α
its␈α
``education,''␈αthe
␈↓ α←␈↓↓␈↓ βWknowledge␈α
base␈α
may,␈α
for␈α
instance,␈α
not␈α
yet␈α
have␈α
information␈α
about␈αall
␈↓ α←␈↓↓␈↓ βWthe␈αgramstain␈αvalues,␈αand␈αthis␈α
might␈αbecome␈αevident␈αin␈αthe␈α
course␈αof
␈↓ α←␈↓↓␈↓ βWteaching␈α∂it␈α∂about␈α⊂a␈α∂new␈α∂organism.␈α∂ Examples␈α⊂of␈α∂this␈α∂are␈α⊂found␈α∂in
␈↓ α←␈↓↓␈↓ βWadditional␈α↔traces␈α↔later␈α↔in␈α↔this␈α↔chapter,␈α↔which␈α↔demonstrate␈α↔that
␈↓ α←␈↓↓␈↓ βW␈↓¬TEIRESIAS␈↓↓ sets up new subtopics as required.}

␈↓ α←␈↓	  What is the morphology of a E.COLI?
␈↓ α←␈↓	  The answer should be in the form a
␈↓ α←␈↓	           <morphology>
␈↓ α←␈↓	  [Type 1 set to a line, then an empty line when done.]
␈↓ α←␈↓	  ++** ␈↓αROD␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓	[Adding E.COLI to ORGANISMS]

␈↓ α←␈↓↓␈↓ βW{In␈αaddition␈α
to␈αcreating␈α
the␈αnew␈αstructure,␈α
it's␈αnecessary␈α
to␈αadd␈α
it␈αto
␈↓ α←␈↓↓␈↓ βWan␈αinternal␈αlist␈α
called␈α␈↓	ORGANISMS␈↓↓.␈αThe␈α
message␈αis␈αprinted␈α
simply␈αto
␈↓ α←␈↓↓␈↓ βWindicate that the proper step has been taken.}

␈↓ α←␈↓	  Ok, done with E.COLI now...
␈↓ α←␈↓	  Back to the rule.

␈↓ α←␈↓	This may take a few moments.

␈↓ α←␈↓↓␈↓ βW{Acquiring the rule then can continue as before.}
␈↓ α←␈↓␈↓140    KNOWLEDGE ACQUISITION II␈↓ 
#6-7␈↓

␈↓"β␈↓ α←␈↓␈↓α6-7-2    Acquisition of a new culture site␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Having␈α∀taught␈α∪the␈α∀performance␈α∀program␈α∪about␈α∀e.coli,␈α∀the␈α∪expert
␈↓ α←␈↓might␈α
later␈α
start␈α
adding␈α
rules␈α
about␈α
the␈α
urinary␈α
tract␈α
and␈α
for␈α
the␈α
first␈α
time
␈↓ α←␈↓mention␈α⊂urine␈α∂as␈α⊂a␈α∂culture␈α⊂site.␈α⊂The␈α∂next␈α⊂example␈α∂shows␈α⊂how␈α⊂this␈α∂would
␈↓ α←␈↓proceed␈αand␈αdemonstrates␈α␈↓¬TEIRESIAS␈↓'s␈αhandling␈αof␈αa␈αfairly␈αcomplex␈αset␈αof␈αdata
␈↓ α←␈↓structure interrelationships.

␈↓ α←␈↓	***************************************************************

␈↓ α←␈↓	The new rule will be called RULE384
␈↓ α←␈↓	  If 1 -␈↓α THERE IS NO HISTORY OF PYELONEPHRITIS␈↓	
␈↓ α←␈↓	     2 -␈↓α THE ORGANISM WAS CULTURED FROM THE URINE␈↓	
␈↓ α←␈↓	     3 -␈↓α THERE IS NO HISTORY OF RECURRENT UTI'S␈↓	
␈↓ α←␈↓	     4 -
␈↓ α←␈↓	 Then
␈↓ α←␈↓	     1 -␈↓α THE ORGANISM IS LIKELY (.3) TO BE E.COLI␈↓	


␈↓ α←␈↓	Sorry, but I don't understand
␈↓ α←␈↓	   THE ORGANISM WAS CULTURED FROM THE URINE
␈↓ α←␈↓	because these words are unknown:  URINE

␈↓ α←␈↓	Would you care to try rephrasing that?
␈↓ α←␈↓	++** ␈↓αY␈↓	
␈↓ α←␈↓	  (Please use *'s to mark what you think might be a new
␈↓ α←␈↓	   attribute or a new value of an attribute)
␈↓ α←␈↓	        3 -␈↓α  THE SITE OF THE CULTURE IS * URINE *␈↓	
␈↓ α←␈↓	It looks as though
␈↓ α←␈↓	     URINE
␈↓ α←␈↓	refers to a new value of an attribute, correct?
␈↓ α←␈↓	++** ␈↓αY␈↓	
␈↓ α←␈↓	 ...is it a new value of the site of a culture?
␈↓ α←␈↓	++** ␈↓αY␈↓	

␈↓ α←␈↓	  Now tell me a few things about it...

␈↓ α←␈↓	  Please give the full, formal name for "URINE"
␈↓ α←␈↓	  ++** ␈↓αURINE␈↓	

␈↓ α←␈↓	  Now please give all synonyms or abbreviations for URINE
␈↓ α←␈↓	  which you would like the system to accept:
␈↓ α←␈↓	  [type an empty line when done]
␈↓ α←␈↓	  ++**

␈↓ α←␈↓	  Please give a short description of URINE as a culture site.
␈↓ α←␈↓	  [type an empty line when done]
␈↓ α←␈↓	  ++** ␈↓αTHERE ARE SEVERAL METHODS OF OBTAINING URINE␈↓	
␈↓ α←␈↓	  ++** ␈↓αSPECIMENS, SOME MORE LIKELY TO PRODUCE STERILE␈↓	
␈↓ α←␈↓	  ++** ␈↓αRESULTS. BECAUSE OF THE LARGE POSSIBILITY OF␈↓	
␈↓ α←␈↓	  ++** ␈↓αCONTAMINATION, CULTURES ARE NOT CONSIDERED␈↓	
␈↓ α←␈↓	  ++** ␈↓αSIGNIFICANT UNLESS COLONY COUNT IS 100,000 OR␈↓	
␈↓ α←␈↓␈↓6-7␈↓ πNACQUIRING NEW VALUES    141␈↓

␈↓"β␈↓ α←␈↓	  ++** ␈↓αMORE.␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓	  Should this new item be added to-
␈↓ α←␈↓	       1 - those sites that are normally sterile, or
␈↓ α←␈↓	       2 - those sites that are normally nonsterile, or
␈↓ α←␈↓	       3 - those sites from which the sterility of cultures
␈↓ α←␈↓	           depends upon the method of collection
␈↓ α←␈↓	  Select one of the above
␈↓ α←␈↓	  ++** ␈↓α3␈↓	

␈↓ α←␈↓	[Adding URINE to OTHERSITES]

␈↓ α←␈↓↓␈↓ βW{Here␈α_we␈α_begin␈α_to␈α_encounter␈α_the␈α_interactions␈α_with␈α_other␈α_data
␈↓ α←␈↓↓␈↓ βWstructures.␈α
 The␈αfirst␈α
problem␈α
is␈αto␈α
find␈αout␈α
whether␈α
the␈αsite␈α
is␈α
to␈αbe
␈↓ α←␈↓↓␈↓ βWconsidered␈α
sterile,␈αnonsterile,␈α
or␈αindeterminant,␈α
so␈αthat␈α
it␈αcan␈α
then␈αbe
␈↓ α←␈↓↓␈↓ βWadded to the proper internal collection of sites.}

␈↓ α←␈↓	[Adding URINE to ALLSITES]

␈↓ α←␈↓↓␈↓ βW{The␈α∪schema␈α∪also␈α∪indicates␈α∪that␈α∪there␈α∪is␈α∪an␈α∪internal␈α∪list␈α∪(called
␈↓ α←␈↓↓␈↓ βWALLSITES)␈α
that␈α
contains␈α
all␈α
the␈αsites.␈α
 The␈α
system␈α
adds␈α
the␈αnew␈α
site
␈↓ α←␈↓↓␈↓ βWto it without having to ask.}

␈↓ α←␈↓	  Please indicate the infections which can occur at the
␈↓ α←␈↓	  site: URINE.  The answer should be in the form of 0 or
␈↓ α←␈↓	  more of the following-
␈↓ α←␈↓	           <an infection name>
␈↓ α←␈↓	  [Type 1 set to a line, then an empty line when done.]
␈↓ α←␈↓	  ++** ␈↓αACUTE-PYELONEPHRITIS␈↓	
␈↓ α←␈↓	  ++** ␈↓αPYELITIS␈↓	
␈↓ α←␈↓	  ++** ␈↓αURETHRITIS␈↓	
␈↓ α←␈↓	  ++** ␈↓αCYSTITIS␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{A␈α
third␈α
data␈αstructure␈α
keeps␈α
track␈α
of␈αwhich␈α
infections␈α
occur␈αat␈α
which
␈↓ α←␈↓↓␈↓ βWsites.}

␈↓ α←␈↓	  Should this new item be added to -
␈↓ α←␈↓	         - those sites for which a positive culture implies a
␈↓ α←␈↓	           specific infection
␈↓ α←␈↓	  ++** ␈↓αN␈↓	

␈↓ α←␈↓↓␈↓ βW{A␈α∂fourth␈α∂contains␈α∂the␈α∞list␈α∂of␈α∂all␈α∂sites␈α∞for␈α∂which␈α∂a␈α∂positive␈α∞culture
␈↓ α←␈↓↓␈↓ βWimplies a specific infection.}

␈↓ α←␈↓	  Please indicate the methods of collection generally
␈↓ α←␈↓	  associated with a culture taken from the URINE, and the
␈↓ α←␈↓	  likelihood that the culture will be sterile.
␈↓ α←␈↓	  The answer should be in the form of 0 or more of the
␈↓ α←␈↓␈↓142    KNOWLEDGE ACQUISITION II␈↓ 
#6-7␈↓

␈↓"β␈↓ α←␈↓	  following-
␈↓ α←␈↓	    <the method of collection of a culture> <certainty factor>
␈↓ α←␈↓	  [Type 1 set to a line, then an empty line when done.]
␈↓ α←␈↓	  ++** ␈↓αVOIDED .7␈↓	
␈↓ α←␈↓	  ++** ␈↓αFOLEY-CATH-ASPIRATION 1.0␈↓	
␈↓ α←␈↓	  ++** ␈↓αFOLEY-BAG-DRAINAGE .9␈↓	
␈↓ α←␈↓	  ++** ␈↓αURETERAL 1.0␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{The␈αfifth␈α
keeps␈αtrack␈α
of␈αthe␈αvarious␈α
ways␈αsamples␈α
can␈αbe␈αcollected␈α
at
␈↓ α←␈↓↓␈↓ βWdifferent sites and notes their likely sterility.}

␈↓ α←␈↓	  What are the likely pathogens to be found at the site: URINE?
␈↓ α←␈↓	  The answer should be in the form of 0 or more of the
␈↓ α←␈↓	  following-
␈↓ α←␈↓	      <identity of an organism> <certainty factor>
␈↓ α←␈↓	  [Type 1 set to a line, then an empty line when done.]
␈↓ α←␈↓	  ++** ␈↓αE.COLI .3␈↓	
␈↓ α←␈↓	  ++** ␈↓αPSEUDOMONAS .2␈↓	
␈↓ α←␈↓	  ++** ␈↓αENTEROCOCCUS .2␈↓	
␈↓ α←␈↓	  ++** ␈↓αPROTEUS .2␈↓	
␈↓ α←␈↓	  ++** ␈↓αKLEBSIELLA .1␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{And␈αthe␈α
last␈αkeeps␈α
track␈αof␈α
which␈αorganisms␈α
are␈αtypically␈α
found␈αat
␈↓ α←␈↓↓␈↓ βWspecific␈α
sites.␈α∞Note␈α
that␈α
all␈α∞of␈α
this␈α
information␈α∞is␈α
required␈α∞for␈α
every
␈↓ α←␈↓↓␈↓ βWnew site that is added to the system.}

␈↓ α←␈↓	  Ok, done with URINE now...
␈↓ α←␈↓	  Back to the rule.

␈↓ α←␈↓↓␈↓ βW{Having␈α
finished␈αwith␈α
the␈αnew␈α
site,␈α␈↓¬TEIRESIAS␈↓↓␈α
returns␈αto␈α
parsing␈αthe
␈↓ α←␈↓↓␈↓ βWnew rule.}

␈↓ α←␈↓	This may take a few moments.
␈↓ α←␈↓␈↓6-8␈↓ ∧SKNOWLEDGE ABOUT REPRESENTATIONS:  ORGANIZATION    143␈↓

␈↓"β␈↓ α←␈↓␈↓α6-8    KNOWLEDGE ABOUT REPRESENTATIONS:  ORGANIZATION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?We␈α∪have␈α∀emphasized␈α∪earlier␈α∪that␈α∀performance␈α∪of␈α∀the␈α∪acquisition
␈↓ α←␈↓system␈α∩(␈↓¬TEIRESIAS␈↓)␈α∩depends␈α∩on␈α∩an␈α⊃extensive␈α∩store␈α∩of␈α∩knowledge␈α∩about␈α⊃the
␈↓ α←␈↓representations␈α
used␈αby␈α
the␈αperformance␈α
program␈α(␈↓¬MYCIN␈↓).␈α
 There␈α
are␈αseveral
␈↓ α←␈↓kinds␈α∃of␈α∃information␈α∀involved␈α∃(Fig.␈α∃6-2␈α∀below).␈α∃ First,␈α∃a␈α∀generalization
␈↓ α←␈↓hierarchy␈α∩links␈α∩the␈α⊃schemata␈α∩together,␈α∩indicating␈α⊃what␈α∩categories␈α∩of␈α⊃data
␈↓ α←␈↓structures␈α∂exist␈α∞in␈α∂the␈α∞system␈α∂and␈α∞the␈α∂relationships␈α∞between␈α∂them.␈α∞ Second,
␈↓ α←␈↓there␈α_are␈α_the␈α_individual␈α_schemata,␈α_each␈α_indicating␈α_the␈α→structure␈α_and
␈↓ α←␈↓interrelationships␈α∂of␈α∂a␈α∂single␈α∂type␈α⊂of␈α∂data␈α∂structure.␈α∂ Finally,␈α∂there␈α⊂are␈α∂the
␈↓ α←␈↓``slotnames''␈α⊂(and␈α⊂associated␈α⊂structures)␈α⊂from␈α⊂which␈α⊂the␈α⊂schemata␈α⊂are␈α∂built;
␈↓ α←␈↓these␈α⊗offer␈α⊗knowledge␈α⊗about␈α⊗specific␈α⊗conventions␈α⊗at␈α⊗the␈α∃programming-
␈↓ α←␈↓language␈α∀level.␈α∀ Each␈α∀of␈α∀these␈α∀supplies␈α∀a␈α∀different␈α∀sort␈α∃of␈α∀information;
␈↓ α←␈↓together␈α∞they␈α∞compose␈α∂an␈α∞extensive␈α∞body␈α∂of␈α∞knowledge␈α∞about␈α∂the␈α∞structure
␈↓ α←␈↓and organization of the representations.


␈↓"β␈↓ α←␈↓	schema hierarchy  ␈↓-- indicates categories of representations and␈↓	
␈↓"β␈↓ α←␈↓	                  ␈↓    interrelations␈↓	
␈↓"β␈↓ α←␈↓	individual schema ␈↓-- describes structure of a single representation␈↓	
␈↓"β␈↓ α←␈↓	slotnames         ␈↓-- the schema building blocks, describe implementation␈↓	
␈↓"β␈↓ α←␈↓	                  ␈↓    conventions


␈↓"β␈↓ α←␈↓α␈↓ βZFig. 6-2.    Types of knowledge about representations.    


␈↓"β␈↓ α←␈↓␈↓α6-8-1    The schema hierarchy␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αschemata␈α
are␈αorganized␈α
into␈αa␈α
generalization␈αhierarchy␈α
that␈αhas
␈↓ α←␈↓several␈αuseful␈αproperties.␈α Part␈α
of␈αthe␈αhierarchy␈αfor␈αthe␈α
current␈αperformance
␈↓ α←␈↓program is shown in the figure below.␈↓
3␈↓
␈↓"β␈↓ α←␈↓␈↓ β?␈↓	KSTRUCT-SCHEMA␈↓␈α(knowledge␈α
structure)␈αsimply␈αprovides␈α
a␈αroot␈αfor␈α
the
␈↓ α←␈↓network;␈α∩its␈α∩schema␈α∩is␈α⊃empty.␈α∩ Below␈α∩it␈α∩are␈α⊃the␈α∩schemata␈α∩for␈α∩value␈α⊃and
␈↓ α←␈↓attribute,␈α↔and␈α↔each␈α↔of␈α↔these␈α↔is␈α↔further␈α↔subdivided␈α↔into␈α_more␈α↔specific
␈↓ α←␈↓schemata.␈α The␈αright␈αbranch␈αof␈αthe␈α
network␈αillustrates␈αthe␈αfact␈αthat␈αa␈α
schema
␈↓ α←␈↓can have more than one parent.








␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[3]␈α∪The␈α∩schemata␈α∪for␈α∩␈↓↓blank␈↓,␈α∪␈↓↓advice␈↓,␈α∩␈↓↓slotname␈↓,␈α∪and␈α∩the␈α∪remainder␈α∪of␈α∩the
␈↓ α←␈↓primitives␈α⊂in␈α⊂Section␈α∂2-4-4␈α⊂each␈α⊂form␈α∂a␈α⊂branch␈α⊂of␈α∂the␈α⊂network␈α⊂one␈α∂level
␈↓ α←␈↓below ␈↓	KSTRUCT-SCHEMA␈↓.  They are omitted here for simplicity.
␈↓ α←␈↓␈↓144    KNOWLEDGE ACQUISITION II␈↓ 
#6-8␈↓


␈↓"β␈↓ α←␈↓	                 KSTRUCT-SCHEMA

␈↓"β␈↓ α←␈↓	   VALUE-SCHEMA                   ATTRIB-SCHEMA


␈↓"β␈↓ α←␈↓	SITE-       IDENT-
␈↓"β␈↓ α←␈↓	SCHEMA      SCHEMA


␈↓"β␈↓ α←␈↓	              PTATTRIB-   INFATTRIB-   CULATTRIB-   ORGATTRIB-
␈↓"β␈↓ α←␈↓	              SCHEMA      SCHEMA       SCHEMA       SCHEMA




␈↓"β␈↓ α←␈↓	                      SVA-SCHEMA  MVA-SCHEMA  TFA-SCHEMA


␈↓"β␈↓ α←␈↓α␈↓ ∧2Fig. 6-3.    Part of the schema hierarchy.    

␈↓"β␈↓ α←␈↓␈↓ β?The␈α_major␈α→contribution␈α_of␈α_the␈α→hierarchy␈α_is␈α_as␈α→an␈α_organizing
␈↓ α←␈↓mechanism␈α
that␈αoffers␈α
a␈α
convenient␈αoverview␈α
of␈α
all␈αthe␈α
representations␈αin␈α
the
␈↓ α←␈↓system.␈α It␈αalso␈αindicates␈α
their␈αglobal␈αorganization.␈α The␈αright␈α
branch␈αabove,
␈↓ α←␈↓for␈α
instance,␈α
indicates␈α
that␈α
there␈α
are␈α
two␈α
different␈α
breakdowns␈α
of␈α
the␈α∞set␈α
of
␈↓ α←␈↓attributes: ␈α↔one␈α↔containing␈α_four␈α↔categories,␈↓
4␈↓␈α↔the␈α↔other␈α_containing␈α↔three
␈↓ α←␈↓categories.␈↓
5␈↓␈αAs␈αwill␈αbe␈αillustrated␈αfurther␈αon,␈αacquisition␈αof␈αa␈αnew␈αinstance␈αof
␈↓ α←␈↓a␈α
conceptual␈αprimitive␈α
is,␈α
in␈αpart,␈α
a␈α
process␈αof␈α
descent␈α
through␈αthis␈α
hierarchy,
␈↓ α←␈↓so it provides a useful structuring of the acquisition dialog.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈αthe␈αacquisition␈αof␈α
new␈αtypes␈αof␈αconceptual␈αprimitives␈α
is␈αviewed
␈↓ α←␈↓as␈α∂a␈α∂process␈α∂of␈α⊂adding␈α∂new␈α∂branches␈α∂to␈α⊂this␈α∂network,␈α∂it␈α∂is␈α⊂important␈α∂that
␈↓ α←␈↓network␈α⊂growth␈α∂be␈α⊂reasonably␈α∂smooth␈α⊂and␈α∂convenient.␈α⊂ Later␈α⊂sections␈α∂will
␈↓ α←␈↓demonstrate␈α⊂that␈α⊂it␈α⊂does,␈α⊂in␈α⊂fact,␈α⊂arise␈α⊂as␈α⊂a␈α⊂natural␈α⊂part␈α⊂of␈α⊃enlarging␈α⊂the
␈↓ α←␈↓knowledge␈αbase␈α
and␈αthat␈αthis␈α
new␈αgrowth␈α
is␈αautomatically␈αreflected␈α
afterward
␈↓ α←␈↓in future dialogs.
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[4]␈α∂The␈α∂attributes␈α∂can␈α∂be␈α∂classified␈α∂according␈α∂to␈α∂which␈α∂object␈α∂they␈α∂are␈α∞an
␈↓ α←␈↓attribute of (e.g., patient, infection, culture, organism).

␈↓"β␈↓ α←␈↓[5]␈α∂They␈α∂can␈α∂also␈α∂be␈α∂broken␈α∂down␈α∂into␈α⊂``single-valued,''␈α∂``multiple-valued,''
␈↓ α←␈↓and␈α∩``true/false''␈α∩types.␈α∩ Single-valued␈α∩attributes␈α∩can␈α∩have␈α∩only␈α∩one␈α⊃value
␈↓ α←␈↓known␈α
with␈α
certainty␈α
(e.g.,␈αan␈α
organism␈α
can␈α
have␈αonly␈α
a␈α
single␈α
identity␈αthat
␈↓ α←␈↓has␈α
a␈α∞CF␈α
of␈α∞1.0),␈α
while␈α
multiple-valued␈α∞attributes␈α
can␈α∞have␈α
more␈α∞than␈α
one
␈↓ α←␈↓(e.g.,␈α⊂there␈α⊂may␈α⊂be␈α∂more␈α⊂than␈α⊂one␈α⊂drug␈α∂to␈α⊂which␈α⊂the␈α⊂patient␈α⊂is␈α∂definitely
␈↓ α←␈↓allergic).␈α
 The␈α
final␈α
category␈αcontains␈α
attributes␈α
that␈α
ask␈α
questions␈αanswered
␈↓ α←␈↓by ``yes'' or ``no'' (e.g., ``Did the organism grow in the aerobic bottle?'').
␈↓ α←␈↓␈↓6-8␈↓ ∧SKNOWLEDGE ABOUT REPRESENTATIONS:  ORGANIZATION    145␈↓

␈↓"β␈↓ α←␈↓␈↓ β?In␈αthe␈αnetwork,␈αextensive␈α
use␈αis␈αmade␈αof␈α
the␈αconcept␈αof␈αinheritance␈α
of
␈↓ α←␈↓properties.␈α⊂ The␈α∂left␈α⊂branch␈α∂above,␈α⊂for␈α∂instance,␈α⊂indicates␈α∂that␈α⊂culture␈α∂site
␈↓ α←␈↓and␈α∂organism␈α∞identity␈α∂are␈α∂more␈α∞specific␈α∂categories␈α∂of␈α∞the␈α∂data␈α∂type␈α∞␈↓	VALUE␈↓.
␈↓ α←␈↓All␈α
of␈α
the␈α
characteristics␈α
that␈α
site␈α
and␈α
identity␈α
have␈α
in␈α
common␈α
as␈α␈↓	VALUE␈↓s␈α
are
␈↓ α←␈↓stored␈α∩in␈α⊃the␈α∩␈↓	VALUE-SCHEMA␈↓.␈α∩ Thus␈α⊃the␈α∩structure␈α⊃description␈α∩part␈α∩of␈α⊃the
␈↓ α←␈↓␈↓	VALUE-SCHEMA␈↓␈α≥(shown␈α≥in␈α≥the␈α≥next␈α≥section)␈α≥describes␈α≥the␈α≤structural
␈↓ α←␈↓components␈α
that␈α
are␈α
common␈α
to␈α∞all␈α
␈↓	VALUE␈↓s.␈α
 The␈α
network␈α
then␈α∞branches␈α
at
␈↓ α←␈↓this␈α
point␈α
because␈α
an␈α
organism␈α
identity␈α
is␈α
a␈α
different␈α
type␈α
of␈α
data␈αstructure
␈↓ α←␈↓than␈α⊂a␈α⊂culture␈α⊂site,␈α⊂and␈α⊂differs␈α⊂in␈α⊂some␈α⊂details␈α⊂of␈α⊂structure.␈α⊂ As␈α⊃the␈α⊂next
␈↓ α←␈↓section␈α∞illustrates,␈α
this␈α∞inheritance␈α∞of␈α
properties␈α∞is␈α∞used␈α
for␈α∞all␈α∞the␈α
different
␈↓ α←␈↓types of information stored in the schema.
␈↓"β␈↓ α←␈↓␈↓ β?This␈αhierarchical␈αdistribution␈αof␈αinformation␈αalso␈αoffers␈αsome␈αhandle
␈↓ α←␈↓on␈α∂the␈α∂issue␈α∂of␈α∂the␈α∂level␈α∂of␈α∂abstraction␈α∂at␈α∂which␈α∂data␈α∂types␈α∂are␈α∞described,
␈↓ α←␈↓since␈α⊂the␈α⊂hierarchy␈α⊂stores␈α⊂at␈α⊂each␈α∂level␈α⊂only␈α⊂those␈α⊂details␈α⊂relevant␈α⊂to␈α∂that
␈↓ α←␈↓particular level.␈↓
6␈↓
␈↓"β␈↓ α←␈↓␈↓ β?While␈α∩it␈α⊃is␈α∩not␈α⊃evident␈α∩from␈α⊃the␈α∩segment␈α⊃of␈α∩the␈α∩schema␈α⊃network
␈↓ α←␈↓shown␈α∞above,␈α∞functions␈α∞constitute␈α∞a␈α∞branch␈α∞of␈α∞the␈α∞network.␈α∞Included␈α∞there,
␈↓ α←␈↓for␈αinstance,␈αare␈αthe␈αpredicate␈αfunctions␈αused␈αin␈αrules.␈α We␈αare␈α
thus␈αviewing
␈↓ α←␈↓functions␈α
as␈α
another␈α
type␈α
of␈αdata␈α
structure.␈α
 Restated␈α
in␈α
␈↓¬LISP␈↓␈α
terms,␈αa␈α
function
␈↓ α←␈↓is␈α``just''␈αanother␈αdata␈αstructure␈αthat␈αhappens␈αto␈αhave␈αan␈αitem␈αon␈αits␈αproperty
␈↓ α←␈↓list␈αcalled␈α␈↓	EXPR␈↓␈α(where␈αthe␈αdefinition␈αis␈αstored).␈α As␈αwill␈αbecome␈αclear,␈αit␈αis␈αat
␈↓ α←␈↓times␈αuseful␈αto␈αtake␈αthis␈αview,␈αbut␈αit␈αis␈αnot␈αin␈αany␈αsense␈αexclusive.␈α Functions
␈↓ α←␈↓will␈α
be␈αviewed␈α
as␈α
both␈αdata␈α
structures␈α
and␈αprocedures,␈α
depending␈α
on␈αwhich
␈↓ α←␈↓is the most relevant at the moment.

␈↓"β␈↓ α←␈↓␈↓α6-8-2    Schema organization␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
schemata␈α
are␈α
the␈α
second␈αof␈α
the␈α
three␈α
kinds␈α
of␈α
knowledge␈αabout
␈↓ α←␈↓representations␈α∂noted␈α∂in␈α∂Fig.␈α∂6-2.␈α∂ Each␈α∂contains␈α∂several␈α∂different␈α∂types␈α∞of
␈↓ α←␈↓information:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?the structure of its instances,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?interrelationships with other data structures,
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?a pointer to all current instances,
␈↓"β␈↓ α←␈↓␈↓ ββ(d)␈↓ β?inter-schema organizational information, and
␈↓"β␈↓ α←␈↓␈↓ ββ(e)␈↓ β?bookkeeping information.

␈↓"β␈↓ α←␈↓␈↓ β?Fig.␈α∩6-4␈α∩shows␈α∩the␈α∩schema␈α∩for␈α⊃the␈α∩value␈α∩of␈α∩an␈α∩attribute␈α∩and␈α⊃the
␈↓ α←␈↓schema␈αfor␈αthe␈αidentity␈αof␈αan␈αorganism.␈α In␈αboth,␈αinformation␈αcorresponding
␈↓ α←␈↓to␈αeach␈α
of␈αthe␈αcategories␈α
listed␈αabove␈α
is␈αgrouped␈αtogether␈α
(the␈αnumbers␈αat␈α
the
␈↓ α←␈↓right are for reference only).
␈↓"β␈↓ α←␈↓␈↓ β?Note␈α∀that,␈α∪since␈α∀the␈α∪␈↓	VALUE-SCHEMA␈↓␈α∀is␈α∪the␈α∀parent␈α∪of␈α∀the␈α∪␈↓	IDENT-
␈↓ α←␈↓	SCHEMA␈↓␈αin␈αthe␈α
hierarchy,␈αinformation␈αin␈αthe␈α
former␈αneed␈αnot␈α
be␈αreproduced

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[6]␈αThe␈αschema␈αhierarchy␈αcan␈αalso␈αbe␈αseen␈αas␈αa␈αdata␈αstructure␈αversion␈αof␈αthe
␈↓ α←␈↓sort of hierarchy often represented with the ␈↓↓class␈↓ construct in ␈↓¬SIMULA␈↓ [Dahl70].
␈↓ α←␈↓␈↓146    KNOWLEDGE ACQUISITION II␈↓ 
#6-8␈↓

␈↓"β␈↓ α←␈↓in␈α∂the␈α∂latter.␈α∂ Hence␈α∂the␈α∂complete␈α∂specification␈α∂for␈α∂an␈α∂organism␈α∂identity␈α∂is
␈↓ α←␈↓given by considering information in both schemata.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈α∞also␈α∞that␈α∞the␈α∞schema␈α∞use␈α∞what␈α∞is␈α∞known␈α∞as␈α∞an␈α
``item-centered''
␈↓ α←␈↓factorization␈α∪and␈α∪indexing␈α∪of␈α∪knowledge.␈α∪ That␈α∪is,␈α∪the␈α∪items␈α∀dealt␈α∪with
␈↓ α←␈↓(␈↓	IDENT␈↓ities,␈α
␈↓	VALUE␈↓s,␈α
etc.)␈α
are␈α
used␈α
as␈αthe␈α
main␈α
indexing␈α
points␈α
for␈α
the␈αbody
␈↓ α←␈↓of␈α∞knowledge␈α∞about␈α∞representations,␈α∂and␈α∞all␈α∞information␈α∞about␈α∂a␈α∞particular
␈↓ α←␈↓item␈α∞is␈α∞associated␈α∞directly␈α∞with␈α
that␈α∞item.␈α∞ The␈α∞advantage␈α∞of␈α∞this␈α
approach
␈↓ α←␈↓lies␈α∂in␈α∂making␈α∂possible␈α∂a␈α∂strongly␈α∞modular␈α∂system␈α∂in␈α∂which␈α∂it␈α∂is␈α∞relatively
␈↓ α←␈↓easy␈αto␈α
organize␈αand␈α
represent␈αa␈α
large␈αbody␈α
of␈αknowledge.␈α
 The␈αtwenty-five
␈↓ α←␈↓or␈α∂so␈α∞schemata␈α∂that␈α∞make␈α∂up␈α∞that␈α∂body␈α∞of␈α∂knowledge␈α∞encode␈α∂a␈α∞significant
␈↓ α←␈↓amount␈α
of␈α
information␈α
about␈α∞the␈α
representation␈α
conventions␈α
of␈α
a␈α∞large␈α
and
␈↓ α←␈↓complex␈α∀program.␈α∃ They␈α∀were␈α∀reasonably␈α∃easy␈α∀to␈α∀construct␈α∃because␈α∀the
␈↓ α←␈↓individual␈α⊂representations␈α⊂are␈α⊂``mostly␈α⊂independent''␈α⊂(i.e.,␈α⊂they␈α⊂have␈α⊂only␈α∂a
␈↓ α←␈↓few,␈α∃well-specified␈α∃kinds␈α∃of␈α∀interactions)␈α∃and␈α∃because␈α∃the␈α∀item-centered
␈↓ α←␈↓organization encourages taking advantage of that modularity.

␈↓"β␈↓ α←␈↓␈↓αInstance structure␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∞part␈α∂of␈α∞the␈α∂schema␈α∞that␈α∂describes␈α∞the␈α∂structure␈α∞of␈α∂its␈α∞instances
␈↓ α←␈↓(lines␈α1-7,␈α15-20)␈αis␈αthe␈αelement␈αthat␈αcorresponds␈αmost␈αclosely␈αto␈αan␈αordinary
␈↓ α←␈↓record␈α∞descriptor.␈α
 The␈α∞current␈α∞implementation␈α
takes␈α∞a␈α
very␈α∞simple␈α∞view␈α
of
␈↓ α←␈↓␈↓¬LISP␈↓␈α∞data␈α∞structures.␈α∞ It␈α
assumes␈α∞that␈α∞they␈α∞are␈α
composed␈α∞of␈α∞a␈α∞print␈α∞name,␈α
a
␈↓ α←␈↓value,␈α∞and␈α∞a␈α∞property␈α∞list,␈α∞with␈α∞the␈α∞usual␈α∞conventions␈α∞for␈α∞each: ␈α∂The␈α∞print
␈↓ α←␈↓name␈α∂is␈α∂a␈α∂single␈α∂identifier␈α∂by␈α⊂which␈α∂the␈α∂object␈α∂is␈α∂named,␈α∂the␈α∂value␈α⊂is␈α∂an
␈↓ α←␈↓atom␈α∂or␈α∂list␈α∂structure,␈α∂and␈α∂the␈α∂property␈α∂list␈α∂is␈α∂composed␈α∂of␈α∂property-value
␈↓ α←␈↓pairs.␈α
 The␈αfirst␈α
three␈α
items␈αin␈α
the␈α
first␈αschema␈α
above␈α
deal␈αwith␈α
each␈αof␈α
these
␈↓ α←␈↓in turn.
␈↓"β␈↓ α←␈↓␈↓ β?Each item is expressed as a triple of the form:

␈↓"β␈↓ α←␈↓	␈↓ ∧o<slotname>  <blank>  <advice>

␈↓ α←␈↓(We␈α∩use␈α∩the␈α∪term␈α∩``slot''␈α∩from␈α∩the␈α∪work␈α∩on␈α∩frames␈α∩[Minsky74]␈α∪since␈α∩the
␈↓ α←␈↓concept␈α
is␈α
similar,␈α∞but␈α
the␈α
schemata␈α
grew␈α∞out␈α
of,␈α
and␈α
are␈α∞fundamentally␈α
an
␈↓ α←␈↓extension␈α
of,␈α
the␈α
idea␈α
of␈α
a␈αrecord␈α
structure).␈α
 For␈α
the␈α
print␈α
name␈α
of␈αany␈α
value
␈↓ α←␈↓of␈α∂an␈α∂attribute,␈α∂then,␈α∂the␈α∂␈↓↓slotname␈↓␈α∂is␈α∂␈↓	PNTNAME␈↓,␈α∂the␈α∂␈↓↓blank␈↓␈α∂is␈α∂␈↓	ATOM␈↓,␈α⊂and␈α∂the
␈↓ α←␈↓␈↓↓advice␈↓ is ␈↓	ASKIT␈↓.␈↓
7␈↓







␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[7]␈α
All␈α
symbols␈αin␈α
the␈α
schemata␈αare␈α
purely␈α
tokens.␈α They␈α
were␈α
chosen␈α
to␈αbe
␈↓ α←␈↓mnemonic,␈α∀but␈α∀no␈α∀significance␈α∀is␈α∪attached␈α∀to␈α∀any␈α∀particular␈α∀name,␈α∪and
␈↓ α←␈↓nothing depends on the use of the particular set of names chosen.
␈↓ α←␈↓␈↓6-8␈↓ ∧SKNOWLEDGE ABOUT REPRESENTATIONS:  ORGANIZATION    147␈↓








␈↓"β␈↓ α←␈↓¬ ␈↓&VALUE-SCHEMA␈↓)αβ
␈↓"β␈↓ α←␈↓¬     PNTNAME       ATOM      ASKIT                                              ␈↓ 
+[1 
␈↓"β␈↓ α←␈↓¬     VAL           PNTNAME   INSLOT                                             ␈↓ 
+[2 
␈↓"β␈↓ α←␈↓¬     PLIST         [(INSTOF  VALUE-SCHEMA                           GIVENIT     ␈↓ 
+[3 
␈↓"β␈↓ α←␈↓¬                     DESCR   STRING                                 ASKIT       ␈↓ 
+[4 
␈↓"β␈↓ α←␈↓¬                     AUTHOR  ATOM                                   FINDIT      ␈↓ 
+[5 
␈↓"β␈↓ α←␈↓¬                     DATE    INTEGER                                CREATEIT)   ␈↓ 
+[6 
␈↓"β␈↓ α←␈↓¬                    CREATEIT]                                                   ␈↓ 
+[7 

␈↓"β␈↓ α←␈↓¬     STRAN         the value of an attribute                                    ␈↓ 
+[8 
␈↓"β␈↓ α←␈↓¬     FATHER        (KSTRUCT-SCHEMA)                                             ␈↓ 
+[9 
␈↓"β␈↓ α←␈↓¬     OFFSPRING     (IDENT-SCHEMA  SITE-SCHEMA)                                  ␈↓ 
+[10

␈↓"β␈↓ α←␈↓¬     DESCR         the VALUE-SCHEMA describes the format for
␈↓"β␈↓ α←␈↓¬                   a value of an attribute                                      ␈↓ 
+[11
␈↓"β␈↓ α←␈↓¬     AUTHOR        DAVIS                                                        ␈↓ 
+[12
␈↓"β␈↓ α←␈↓¬     DATE          1115                                                         ␈↓ 
+[13
␈↓"β␈↓ α←␈↓¬     INSTOF        (SCHEMA-SCHEMA)                                              ␈↓ 
+[14




␈↓"β␈↓ α←␈↓¬ ␈↓&IDENT-SCHEMA␈↓)αβ
␈↓"β␈↓ α←␈↓¬     PLIST         [(INSTOF  IDENT-SCHEMA                           GIVENIT     ␈↓ 
+[15
␈↓"β␈↓ α←␈↓¬                     SYNONYM (KLEENE (1 0) < ATOM >)                ASKIT       ␈↓ 
+[16
␈↓"β␈↓ α←␈↓¬                     AIR     (KLEENE (1 1 2) <(AIR-INST CF-INST)> ) ASKIT       ␈↓ 
+[17
␈↓"β␈↓ α←␈↓¬                     GRAM    GRAM-INST                              ASKIT       ␈↓ 
+[18
␈↓"β␈↓ α←␈↓¬                     MORPH   MORPH-INST                             ASKIT       ␈↓ 
+[19
␈↓"β␈↓ α←␈↓¬                    CREATEIT]                                                   ␈↓ 
+[20

␈↓"β␈↓ α←␈↓¬     RELATIONS     ((ADDTO (AND* ORGANISMS)))                                   ␈↓ 
+[21

␈↓"β␈↓ α←␈↓¬     INSTANCES     (ACINETOBACTER ACTINOMYCETES  ...  XANTHOMONAS YERSINA)      ␈↓ 
+[22

␈↓"β␈↓ α←␈↓¬     STRAN         the identity of an organism                                  ␈↓ 
+[23
␈↓"β␈↓ α←␈↓¬     FATHER        (VALUE-SCHEMA)                                               ␈↓ 
+[24
␈↓"β␈↓ α←␈↓¬     OFFSPRING     NIL                                                          ␈↓ 
+[25

␈↓"β␈↓ α←␈↓¬     DESCR         the IDENT-SCHEMA describes the format for an organism        ␈↓ 
+[26
␈↓"β␈↓ α←␈↓¬     AUTHOR        DAVIS                                                        ␈↓ 
+[27
␈↓"β␈↓ α←␈↓¬     DATE          1115                                                         ␈↓ 
+[28
␈↓"β␈↓ α←␈↓¬     INSTOF        (SCHEMA-SCHEMA)                                              ␈↓ 
+[29

␈↓"β
␈↓"β
␈↓"β␈↓ α←␈↓α␈↓ ¬∀Fig. 6-4.    Two schemata.    
␈↓ α←␈↓␈↓148    KNOWLEDGE ACQUISITION II␈↓ 
#6-8␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α∪␈↓↓slotname␈↓␈α∪labels␈α∩the␈α∪``kind''␈α∪of␈α∪thing␈α∩that␈α∪fills␈α∪the␈α∪␈↓↓blank␈↓␈α∩and
␈↓ α←␈↓provides␈α∩access␈α⊃to␈α∩other␈α⊃information␈α∩that␈α⊃aids␈α∩in␈α⊃the␈α∩knowledge␈α⊃transfer
␈↓ α←␈↓process.␈α% Slotnames␈α$are␈α%the␈α$conceptual␈α%primitives␈α%around␈α$which
␈↓ α←␈↓representation-specific␈α↔and␈α⊗representation-independent␈α↔knowledge␈α↔in␈α⊗the
␈↓ α←␈↓system␈α
is␈α
organized.␈α
 All␈α
of␈α
the␈α
semantics␈α
of␈α
a␈α
print␈α
name,␈α
for␈α∞instance,␈α
are
␈↓ α←␈↓contained␈αin␈αthe␈α␈↓	PNTNAME␈↓␈αslot␈α
and␈αthe␈αstructures␈αassociated␈αwith␈αit␈α
(described
␈↓ α←␈↓in Section 6-8-3).
␈↓"β␈↓ α←␈↓␈↓ β?The␈α␈↓↓blank␈↓␈αspecifies␈αthe␈α
exact␈αformat␈αof␈αthe␈αinformation␈α
required.␈α A
␈↓ α←␈↓translated␈α∂form␈α∂of␈α∂it␈α∂is␈α∂printed␈α∂out␈α∂when␈α∂requesting␈α∂information␈α∂from␈α∞the
␈↓ α←␈↓expert␈α
and␈αis␈α
then␈α
used␈αto␈α
parse␈αhis␈α
response␈α
and␈αinsure␈α
its␈αsyntactic␈α
validity.
␈↓ α←␈↓The␈α
blank␈α
has␈α
a␈α
simple␈α
syntax␈α
but␈α
can␈α
express␈α
a␈α
range␈α
of␈α∞structures.␈α
 The
␈↓ α←␈↓term␈α∩␈↓	KLEENE␈↓,␈α∩for␈α∩instance,␈α∩is␈α∪taken␈α∩from␈α∩the␈α∩Kleene␈α∩star␈α∩and␈α∪implies␈α∩a
␈↓ α←␈↓repetition␈α
of␈α
the␈αform␈α
within␈α
the␈αangle␈α
brackets.␈α
 The␈αparenthesized␈α
numbers
␈↓ α←␈↓that␈α∀follow␈α∪it␈α∀indicate␈α∪the␈α∀typical,␈α∪minimum,␈α∀and␈α∪maximum␈α∀number␈α∪of
␈↓ α←␈↓occurrences␈α
of␈αthe␈α
form.␈αThe␈α
appearance␈α
of␈αa␈α
term␈αof␈α
the␈αform␈α
␈↓	<datatype>-
␈↓ α←␈↓	INST␈↓ indicates some instance of the ␈↓	<datatype>-SCHEMA␈↓.  Thus,

␈↓"β␈↓ α←␈↓	␈↓ ∧/(KLEENE (1 1 2) <(AIR-INST CF-INST)>)

␈↓ α←␈↓from␈αthe␈αidentity␈α
schema␈αabove␈αindicates␈αthat␈α
the␈αaerobicity␈αof␈α
an␈αorganism
␈↓ α←␈↓is␈α∀described␈α∀by␈α∀1␈α∀or␈α∀2␈α∀lists␈α∀of␈α∀the␈α∀form␈α∀␈↓	(<aerobicity>␈α∪<certainty-
␈↓ α←␈↓	factor>)␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂␈↓↓advice␈↓␈α∂suggests␈α∂how␈α∂to␈α∂find␈α∂the␈α∂information.␈α∂ Various␈α∂sorts␈α∂of
␈↓ α←␈↓information␈αare␈αemployed␈αin␈αthe␈αcourse␈αof␈αacquiring␈αa␈αnew␈αconcept␈αfrom␈αthe
␈↓ α←␈↓expert.␈α Some␈αof␈αit␈αis␈αdomain␈αspecific␈α(e.g.,␈αthe␈αgramstain␈αof␈αa␈αnew␈αorganism)
␈↓ α←␈↓and␈α⊂clearly␈α∂must␈α⊂be␈α∂supplied␈α⊂by␈α∂the␈α⊂expert.␈α∂ Other␈α⊂parts␈α∂of␈α⊂it␈α⊂are␈α∂purely
␈↓ α←␈↓representation␈α∂specific.␈α∂ These␈α∂should␈α⊂be␈α∂supplied␈α∂by␈α∂the␈α∂system␈α⊂itself,␈α∂not
␈↓ α←␈↓only␈α∂because␈α⊂they␈α∂deal␈α⊂with␈α∂information␈α⊂that␈α∂the␈α⊂system␈α∂already␈α⊂has␈α∂(and
␈↓ α←␈↓therefore␈αshould␈α
not␈αhave␈αto␈α
ask),␈αbut␈α
because␈αthe␈αexpert␈α
is␈αassumed␈αto␈α
know
␈↓ α←␈↓nothing␈α⊃about␈α⊃programming.␈α⊃ Even␈α⊃a␈α⊃trivial␈α⊃question␈α⊃concerning␈α⊃internal
␈↓ α←␈↓data␈α∪structure␈α∪management␈α∪would␈α∩thus␈α∪appear␈α∪incomprehensible␈α∪to␈α∩him.
␈↓ α←␈↓The␈α␈↓↓advice␈↓␈αprovides␈αa␈αway␈αof␈αexpressing␈αinstructions␈αto␈αthe␈αsystem␈αon␈αwhere
␈↓ α←␈↓to␈α
find␈α
the␈α
information␈α
it␈α
needs.␈α There␈α
are␈α
five␈α
such␈α
instructions␈α
that␈αcan␈α
be
␈↓ α←␈↓given.

␈↓"β␈↓ α←␈↓	      ASKIT      ␈↓ask the expert␈↓	
␈↓"β␈↓ α←␈↓	      CREATEIT   ␈↓manufacture the answer␈↓	
␈↓"β␈↓ α←␈↓	      FINDIT     ␈↓the answer is available internally, retrieve it␈↓	
␈↓"β␈↓ α←␈↓	      GIVENIT    ␈↓use the contents of the blank as is (like ␈↓¬QUOTE␈↓ in ␈↓¬LISP␈↓)␈↓	
␈↓"β␈↓ α←␈↓	      INSLOT     ␈↓use the contents of the slot indicated

␈↓"β␈↓ α←␈↓␈↓ β?The␈αfirst␈αtriple␈αin␈αFig.␈α6-4␈α(line␈α1)␈αindicates␈αthen␈αthat␈αthe␈αprint␈αname
␈↓ α←␈↓is␈α
an␈α
atom␈α
and␈αthat␈α
it␈α
should␈α
be␈α
requested␈αfrom␈α
the␈α
expert.␈α
 The␈αsecond␈α
(line
␈↓ α←␈↓2)␈α
indicates␈α∞that␈α
the␈α
organism␈α∞name␈α
should␈α
evaluate␈α∞to␈α
its␈α
print␈α∞name,␈α
and
␈↓ α←␈↓the␈α
third␈α(lines␈α
3␈α
-␈α7)␈α
indicates␈αthe␈α
form␈α
of␈αthe␈α
property␈αlist.␈α
 Note␈α
that␈αthe
␈↓ α←␈↓␈↓↓blank␈↓␈αfor␈αthe␈αlast␈αof␈αthese␈αconsists␈αof,␈αin␈αturn,␈αa␈αset␈α
of␈α␈↓↓slotname-blank-advice␈↓
␈↓ α←␈↓triples describing the property list.
␈↓ α←␈↓␈↓6-8␈↓ ∧SKNOWLEDGE ABOUT REPRESENTATIONS:  ORGANIZATION    149␈↓

␈↓"β␈↓ α←␈↓␈↓αInterrelationships␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?A␈α
second␈α
main␈α
function␈αof␈α
the␈α
schema␈α
is␈αto␈α
provide␈α
a␈α
record␈α
of␈αthe
␈↓ α←␈↓interrelationships␈α∞(line␈α
21)␈α∞of␈α
data␈α∞structures.␈α
 The␈α∞␈↓	RELATIONS␈↓␈α∞slot␈α
contains
␈↓ α←␈↓this␈α
information,␈α
expressed␈α
in␈αa␈α
simple␈α
language␈α
for␈α
describing␈αdata␈α
structure
␈↓ α←␈↓relationships.  In BNF terms, it looks like:

␈↓"β␈↓ α←␈↓	   <update>    =  ( <command> ( <switch> <structure>␈↓
+␈↓	)␈↓
+␈↓	)
␈↓"β␈↓ α←␈↓	   <command>   =  ADDTO | EDITFN
␈↓"β␈↓ α←␈↓	   <switch>    =  AND* | OR*  | XOR* | (<switch> <structure>␈↓
+␈↓	)
␈↓"β␈↓ α←␈↓	   <structure> =  <any data structure or function name>

␈↓ α←␈↓(The␈αsuperscript␈α``+''␈αmeans␈α``one␈αor␈αmore.'') ␈α␈↓	ADDTO␈↓␈αindicates␈αthat␈αsome␈αother
␈↓ α←␈↓structure␈αin␈αthe␈αsystem␈αshould␈αbe␈αtold␈αabout␈αthe␈αnew␈αinstance,␈α
while␈α␈↓	EDITFN␈↓
␈↓ α←␈↓indicates␈αthat␈αsome␈αfunction␈α
may␈αneed␈αto␈αbe␈αedited␈α
as␈αa␈αresult␈αof␈αcreating␈α
the
␈↓ α←␈↓new␈α∀instance.␈α∀ The␈α∃three␈α∀switches␈α∀indicate␈α∃that␈α∀the␈α∀action␈α∃specified␈α∀by
␈↓ α←␈↓<command>␈α⊂should␈α⊂be␈α⊂taken␈α⊂on␈α⊂all␈α⊂(␈↓	AND*␈↓),␈α⊂1␈α⊂or␈α⊂more␈α⊂(␈↓	OR*␈↓),␈α⊂or␈α⊂exactly␈α∂1
␈↓ α←␈↓(␈↓	XOR*␈↓)␈α⊃of␈α⊃the␈α⊂structures␈α⊃that␈α⊃follow.␈α⊂ In␈α⊃the␈α⊃case␈α⊂of␈α⊃a␈α⊃new␈α⊃organism,␈α⊂the
␈↓ α←␈↓update␈α∀is␈α∀a␈α∀simple␈α∪one,␈α∀and␈α∀its␈α∀name␈α∪is␈α∀added␈α∀to␈α∀the␈α∀structure␈α∪called
␈↓ α←␈↓␈↓	ORGANISMS␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α$recursive␈α$definition␈α#allows␈α$construction␈α$of␈α#conditional
␈↓ α←␈↓expressions,␈α∂as␈α∂in␈α∂the␈α∂␈↓	RELATIONS␈↓␈α⊂information␈α∂in␈α∂the␈α∂schema␈α∂for␈α⊂a␈α∂culture
␈↓ α←␈↓site:

␈↓"β␈↓ α←␈↓	   ((ADDTO (XOR* STERILESITES NONSTERILESITES OTHERSITES))
␈↓"β␈↓ α←␈↓	    (ADDTO (AND* ALLSITES SITE-INFECT))
␈↓"β␈↓ α←␈↓	    (ADDTO (OR*  PATHOGNOMONIC-SITES))
␈↓"β␈↓ α←␈↓	    (ADDTO ((OR* NONSTERILESITES OTHERSITES) PATH-FLORA))
␈↓"β␈↓ α←␈↓	    (ADDTO ((AND* OTHERSITES) METHOD))))

␈↓ α←␈↓Here,␈α∞the␈α
first␈α∞three␈α
tasks␈α∞are␈α
straightforward,␈α∞but␈α
the␈α∞fourth␈α∞line␈α
indicates
␈↓ α←␈↓that␈αif␈αthe␈αsite␈α
is␈αeither␈αnonsterile␈αor␈α
indeterminant␈αthen␈αit␈αshould␈α
be␈αadded
␈↓ α←␈↓to␈α→the␈α_structure␈α→called␈α_␈↓	PATH-FLORA␈↓.␈α→ The␈α_last␈α→line␈α_indicates␈α→that␈α_all
␈↓ α←␈↓indeterminant sites should be added to the structure called ␈↓	METHOD␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
key␈α
point␈α
here␈α
is␈α
to␈α
provide␈α
the␈α
system␈α
architect␈α
with␈α
a␈α
way␈α
of
␈↓ α←␈↓making␈α∂explicit␈α∂all␈α∞of␈α∂the␈α∂data␈α∞structure␈α∂interrelationships␈α∂upon␈α∂which␈α∞his
␈↓ α←␈↓design␈α∂depends.␈α∂ The␈α⊂approach␈α∂we␈α∂use␈α∂differs␈α⊂slightly␈α∂from␈α∂the␈α⊂one␈α∂more
␈↓ α←␈↓typically␈α∞taken,␈α∂which␈α∞relies␈α∂on␈α∞a␈α∂demon-like␈α∞mechanism␈α∂that␈α∞uses␈α∂the␈α∞full
␈↓ α←␈↓power␈α
of␈α
the␈α
underlying␈α
programming␈αlanguage.␈α
 We␈α
have␈α
avoided␈α
the␈αuse
␈↓ α←␈↓of␈α⊂an␈α⊂arbitrary␈α⊃body␈α⊂of␈α⊂code␈α⊃and␈α⊂emphasized␈α⊂instead␈α⊃the␈α⊂use␈α⊂of␈α⊃a␈α⊂task-
␈↓ α←␈↓specific high-level language.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α.formalization␈α.of␈α.knowledge␈α.about␈α/data␈α.structure
␈↓ α←␈↓interrelationships␈α∩has␈α⊃several␈α∩useful␈α⊃applications.␈α∩ First,␈α⊃since␈α∩the␈α⊃domain
␈↓ α←␈↓expert␈α⊂cannot,␈α⊃in␈α⊂general,␈α⊃be␈α⊂expected␈α⊂to␈α⊃know␈α⊂about␈α⊃such␈α⊂representation
␈↓ α←␈↓conventions,␈α∂expressing␈α∂them␈α⊂in␈α∂machine-accessible␈α∂form␈α∂makes␈α⊂it␈α∂possible
␈↓ α←␈↓for␈α⊃␈↓¬TEIRESIAS␈↓␈α⊃to␈α⊃take␈α⊃over␈α⊃the␈α⊃task␈α⊃of␈α⊃maintaining␈α⊃them.␈α⊃ Second,␈α⊃having
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓␈α
attend␈α
to␈α∞them␈α
insures␈α
a␈α
level␈α∞of␈α
knowledge␈α
base␈α∞integrity␈α
without
␈↓ α←␈↓making␈αunreasonable␈αdemands␈αon␈αthe␈αexpert.␈α Finally,␈αit␈αkeeps␈αknowledge␈αin
␈↓ α←␈↓␈↓150    KNOWLEDGE ACQUISITION II␈↓ 
#6-8␈↓

␈↓"β␈↓ α←␈↓the␈α_system␈α→accessible␈α_since␈α→the␈α_␈↓	RELATIONS␈↓␈α_make␈α→explicit␈α_the␈α→sort␈α_of
␈↓ α←␈↓knowledge␈α
that␈α
is␈α
often␈α
left␈α
implicit,␈αor␈α
which␈α
is␈α
embedded␈α
in␈α
code␈αand␈α
hence
␈↓ α←␈↓is␈α→inaccessible.␈α→ There␈α→are␈α~several␈α→advantages␈α→to␈α→this␈α~accessibility␈α→of
␈↓ α←␈↓knowledge.␈α For␈αexample,␈αby␈αadding␈αto␈α␈↓¬TEIRESIAS␈↓␈αa␈αsimple␈αanalyzer␈αthat␈αcould
␈↓ α←␈↓``read''␈α
the␈α∞␈↓	RELATIONS␈↓,␈α
a␈α
programmer␈α∞could␈α
ask␈α
questions␈α∞like␈α
␈↓↓What␈α∞else␈α
in
␈↓ α←␈↓↓the␈α
system␈α
will␈α
be␈α
affected␈α
if␈α
I␈αadd␈α
a␈α
new␈α
instance␈α
of␈α
this␈α
data␈α
structure?␈↓␈αor
␈↓ α←␈↓␈↓↓What␈αare␈α
all␈αthe␈αother␈α
structures␈αthat␈α
are␈αrelated␈αto␈α
this␈αone?␈↓␈α
This␈αwould␈αbe␈α
a
␈↓ α←␈↓useful form of on-line documentation.␈↓
8␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈α
are␈αadditional␈α
advantages␈αthat␈α
will␈αbecome␈α
apparent␈αin␈α
Section
␈↓ α←␈↓6-9-1, which describes how the updating is actually performed.

␈↓"β␈↓ α←␈↓␈↓αCurrent instances␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Each␈α∩schema␈α∩keeps␈α∪track␈α∩of␈α∩all␈α∩of␈α∪its␈α∩current␈α∩instances␈α∪(line␈α∩22),
␈↓ α←␈↓primarily␈αfor␈αuse␈αin␈αknowledge␈αbase␈αmaintenance.␈α If␈αit␈αbecomes␈αnecessary␈αto
␈↓ α←␈↓make␈α∞changes␈α∞to␈α∂the␈α∞design␈α∞of␈α∞a␈α∂particular␈α∞representation,␈α∞for␈α∂instance,␈α∞we
␈↓ α←␈↓want␈α
to␈α
be␈α
sure␈α
that␈α
all␈α
instances␈α
of␈α
it␈α
are␈α
modified␈α
appropriately.␈α Keeping␈α
a
␈↓ α←␈↓list of all such instances is an obvious but very useful solution.

␈↓"β␈↓ α←␈↓␈↓αOrganizational information␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?␈↓	FATHER␈↓␈α∞indicates␈α∞the␈α∞(more␈α∂general)␈α∞ancestors␈α∞of␈α∞this␈α∞schema␈α∂in␈α∞the
␈↓ α←␈↓hierarchy␈α∂and␈α∂␈↓	OFFSPRING␈↓␈α∂indicates␈α∂its␈α∂more␈α∂specific␈α∂offspring␈α∂(lines␈α∞9-10).
␈↓ α←␈↓␈↓	STRAN␈↓␈α(line␈α8)␈αis␈αan␈αEnglish␈αphrase␈αindicating␈αwhat␈αsort␈αof␈αthing␈αthe␈αschema
␈↓ α←␈↓describes and is used in communicating with the expert.

␈↓"β␈↓ α←␈↓␈↓αBookkeeping information␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Much␈α∪the␈α∀same␈α∪sort␈α∪of␈α∀bookkeeping␈α∪information␈α∪(lines␈α∀11-14)␈α∪is
␈↓ α←␈↓maintained␈αfor␈αeach␈αdata␈αstructure␈αas␈αis␈αkept␈αfor␈αrules;␈α␈↓	DESCR␈↓iption,␈α␈↓	AUTHOR␈↓,
␈↓ α←␈↓and␈α␈↓	DATE␈↓␈α
are␈αthe␈αanalogous␈α
items.␈α␈↓	INSTOF␈↓␈αis␈α
the␈αinverse␈αof␈α
␈↓	INSTANCES␈↓␈αand
␈↓ α←␈↓indicates which schema was used to create this data structure.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈αthat␈αin␈αthe␈αcurrent␈αexample␈αit␈αis␈αthe␈αorganism␈αschema␈αitself␈αthat
␈↓ α←␈↓is␈α
being␈α
described␈α
by␈α
all␈α
of␈αthis␈α
bookkeeping␈α
information,␈α
and,␈α
as␈α
shown,␈αit␈α
is
␈↓ α←␈↓an instance of the ␈↓	SCHEMA-SCHEMA␈↓ (described in Section 6-11).










␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[8]␈α⊂This␈α⊃is␈α⊂the␈α⊃data␈α⊂structure␈α⊂analogue␈α⊃for␈α⊂the␈α⊃facility␈α⊂of␈α⊃␈↓¬INTERLISP␈↓␈α⊂called
␈↓ α←␈↓MASTERSCOPE,␈α∀which␈α∀can␈α∀analyze␈α∪a␈α∀set␈α∀of␈α∀function␈α∀definitions␈α∪and
␈↓ α←␈↓answer questions like ␈↓↓Who calls function F?, Which function binds X?␈↓, etc.
␈↓ α←␈↓␈↓6-8␈↓ ∧SKNOWLEDGE ABOUT REPRESENTATIONS:  ORGANIZATION    151␈↓

␈↓"β␈↓ α←␈↓␈↓α6-8-3    Slotnames and slotexperts␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂most␈α∂detailed␈α∂knowledge␈α∂about␈α∂representations␈α∂is␈α∂found␈α∂in␈α∞the
␈↓ α←␈↓slotnames␈αand␈αthe␈αstructures␈αassociated␈αwith␈αthem.␈αThey␈αdeal␈αwith␈αaspects␈αof
␈↓ α←␈↓the␈α
representation␈α
that␈α
are␈α
at␈α
the␈α
level␈α
of␈α
programming-language␈α
constructs
␈↓ α←␈↓and conventions.  The overall structure of a slotname is shown below.

␈↓"β␈↓ α←␈↓	    ␈↓&<slotname>␈↓)αβ

␈↓"β␈↓ α←␈↓	  PROMPT  ␈↓an English phrase used to request the information to fill the slot␈↓	
␈↓"β␈↓ α←␈↓	  TRANS   ␈↓an English phrase used when displaying the information␈↓	
␈↓"β␈↓ α←␈↓	          ␈↓  found in the slot␈↓	
␈↓"β␈↓ α←␈↓	  EXPERT  ␈↓the name of the slotexpert␈↓	


␈↓"β␈↓ α←␈↓α␈↓ βjFig. 6-5.    Information associated with a slotname.    

␈↓"β␈↓ α←␈↓␈↓ β?The␈α␈↓	PROMPT␈↓␈αand␈α␈↓	TRANS␈↓␈αare␈α
part␈αof␈αthe␈αsimple␈αmechanism␈αthat␈α
makes
␈↓ α←␈↓the␈αcreation␈αof␈αa␈αnew␈αdata␈αstructure␈αan␈αinteractive␈αoperation.␈α The␈αformer␈αis
␈↓ α←␈↓used␈αto␈αrequest␈αinformation,␈αthe␈αlatter␈αis␈αused␈αwhen␈αit␈αis␈αnecessary␈αto␈αdisplay
␈↓ α←␈↓information that has previously been deposited in a slot.␈↓
9␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Associated␈α∞with␈α∞each␈α∞slotname␈α∞is␈α∞a␈α∞procedure␈α∞called␈α∞a␈α∞␈↓↓slotexpert␈↓␈α∞(or
␈↓ α←␈↓simply,␈α∪␈↓↓expert␈↓).␈α∀ It␈α∪serves␈α∀primarily␈α∪as␈α∀a␈α∪repository␈α∀for␈α∪useful␈α∀pieces␈α∪of
␈↓ α←␈↓knowledge␈α→concerning␈α→the␈α→implementation␈α→of␈α→the␈α→representations.␈α→ For
␈↓ α←␈↓example,␈α
names␈α
of␈α
data␈α
structures␈α
have␈α
to␈α
be␈α
unique␈α
to␈α
avoid␈α
confusion␈αor
␈↓ α←␈↓inadvertent␈αmangling.␈α Yet,␈αin␈α
knowledge␈αacquisition,␈αnew␈αdata␈αstructures␈α
are
␈↓ α←␈↓constantly␈α∂being␈α∂created␈α∂and␈α∞many␈α∂of␈α∂their␈α∂names␈α∞are␈α∂chosen␈α∂by␈α∂the␈α∞user.
␈↓ α←␈↓Part␈αof␈αthe␈αtask␈αcarried␈αout␈αby␈αthe␈α␈↓↓expert␈↓␈αassociated␈αwith␈αthe␈α␈↓	PNTNAME␈↓␈αslot␈αis
␈↓ α←␈↓to assure this uniqueness.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂slotexperts␈α⊂are␈α⊂organized␈α⊂around␈α⊂the␈α⊂different␈α⊂sorts␈α⊃of␈α⊂advice
␈↓ α←␈↓that␈α
can␈α
be␈α
used␈α
in␈α
a␈α
slot.␈αTheir␈α
general␈α
format␈α
is␈α
shown␈α
below.␈α
 Since␈αnot
␈↓ α←␈↓all␈α∞pieces␈α∞of␈α∞advice␈α∞are␈α∞meaningful␈α∞for␈α∞all␈α∞slotexperts,␈α∞in␈α∞general␈α∞not␈α∞every
␈↓ α←␈↓slotexpert has an entry for every piece of advice.













␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[9]␈α_The␈α_idea␈α_of␈α↔a␈α_␈↓	PROMPT␈↓␈α_and␈α_␈↓	TRANS␈↓␈α↔were␈α_adapted␈α_from␈α_work␈α↔in
␈↓ α←␈↓[Shortliffe76].
␈↓"β␈↓ α←␈↓␈↓152    KNOWLEDGE ACQUISITION II␈↓ 
#6-8␈↓


␈↓"β␈↓ α←␈↓	 (<slotexpert> [LAMBDA (BLANK ADVICE)
␈↓"β␈↓ α←␈↓	                       (SELECTQ ADVICE
␈↓"β␈↓ α←␈↓	                               (ASKIT    ␈↓∧###␈↓	)
␈↓"β␈↓ α←␈↓	                               (CREATEIT ␈↓∧###␈↓	)
␈↓"β␈↓ α←␈↓	                               (FINDIT   ␈↓∧###␈↓	)
␈↓"β␈↓ α←␈↓	                               (INSLOT   ␈↓∧###␈↓	)
␈↓"β␈↓ α←␈↓	                               (GIVENIT  ␈↓∧###␈↓	)    ␈↓etc.␈↓	])

␈↓"β
␈↓"β␈↓ α←␈↓	␈↓ β'␈↓αFig.␈α_6-6.    The␈α_structure␈α_of␈α_a␈α_slotexpert.  ␈↓␈↓	SELECTQ␈↓␈α_can␈α↔be
␈↓ α←␈↓␈↓ β'thought␈α
of␈αas␈α
a␈α␈↓↓case␈↓␈α
statement␈α
for␈αsymbolic␈α
computation.␈αThus␈α
the
␈↓ α←␈↓␈↓ β'code␈α∞above␈α∞is␈α∞equivalent␈α
to␈α∞␈↓↓if␈α∞ADVICE␈α∞=␈α
ASKIT␈α∞then␈α∞...␈α∞else␈α
if
␈↓ α←␈↓↓␈↓ β'ADVICE = CREATEIT then ...␈↓ etc.

␈↓"β␈↓ α←␈↓␈↓ β?The␈αindividual␈αchunks␈αof␈αcode␈αthat␈αmake␈αup␈αthe␈αparts␈αof␈αthe␈α␈↓↓expert␈↓s
␈↓ α←␈↓are␈α∂the␈α∂smallest␈α∂units␈α∂of␈α∂knowledge␈α∂organization␈α∂in␈α∂our␈α⊂framework.␈α∂ They
␈↓ α←␈↓embody␈α
knowledge␈αabout␈α
things␈α
like␈αwhere␈α
to␈αfind␈α
or␈α
how␈αto␈α
create␈αthe␈α
items
␈↓ α←␈↓needed␈αto␈αfill␈αthe␈α␈↓↓blank␈↓␈αfor␈αa␈αparticular␈αslot.␈α For␈αinstance,␈αwe␈αnoted␈αthat␈αthe
␈↓ α←␈↓␈↓↓expert␈↓␈αassociated␈αwith␈αthe␈α␈↓	PNTNAME␈↓␈αslot␈αinsures␈αthe␈αuniqueness␈αof␈αnames␈αthat
␈↓ α←␈↓are␈αsupplied␈αby␈αthe␈αuser.␈α This␈αroutine␈αwould␈αbe␈αfound␈αin␈αthe␈α␈↓	ASKIT␈↓␈αsection
␈↓ α←␈↓of␈α
the␈α␈↓↓expert␈↓.␈α
 Code␈α
in␈αthe␈α
␈↓	CREATEIT␈↓␈αsection␈α
uses␈α
a␈αnumber␈α
of␈αheuristics␈α
that
␈↓ α←␈↓help␈αto␈αgenerate␈αprint␈αnames␈αthat␈αare␈αbetween␈α4␈αand␈α10␈αcharacters␈αlong␈αand
␈↓ α←␈↓that␈α
are␈α
reasonably␈αmnemonic.␈α
This␈α
is␈αused␈α
when␈α
the␈αsystem␈α
itself␈α
creates␈αa
␈↓ α←␈↓name for a new internal data structure.␈↓
10␈↓

␈↓"β␈↓ α←␈↓␈↓ β?Recall␈α∂that␈α∂we␈α∂set␈α∂out␈α⊂to␈α∂describe␈α∂representations␈α∂in␈α∂order␈α⊂to␈α∂make
␈↓ α←␈↓possible␈α↔the␈α⊗interactive␈α↔acquisition␈α⊗of␈α↔new␈α⊗conceptual␈α↔primitives.␈α⊗ The
␈↓ α←␈↓slotname␈α
and␈α
associated␈α
expert␈α
organize␈α
the␈α
knowledge␈α
needed␈α
and␈αprovide
␈↓ α←␈↓the␈α∀English␈α∪to␈α∀make␈α∪the␈α∀operation␈α∪interactive.␈α∀ The␈α∪blank␈α∀provides␈α∪an
␈↓ α←␈↓indication␈α∂of␈α∞the␈α∂format␈α∂of␈α∞the␈α∂answers␈α∞to␈α∂questions␈α∂and␈α∞a␈α∂check␈α∂on␈α∞their
␈↓ α←␈↓syntax.␈αThe␈αadvice␈αallows␈αthe␈αembedding␈αof␈αan␈αadditional␈αsort␈αof␈αknowledge
␈↓ α←␈↓that makes the process function efficiently and ``intelligently.''

␈↓"β␈↓ α←␈↓␈↓αSlotnames as data structures, ``circularity'' of the formalism␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?While␈α∪discussing␈α∩the␈α∪use␈α∩of␈α∪a␈α∩typed␈α∪language,␈α∩it␈α∪was␈α∪noted␈α∩that

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[10]␈α
While␈α
the␈α
slotnames␈α
are␈α
currently␈α
globally␈α
unique␈α
(the␈α
␈↓	SYNONYM␈↓␈α
slot␈αin
␈↓ α←␈↓Fig.␈α∞6-4,␈α∞for␈α∞instance,␈α∞is␈α∞presumed␈α∞to␈α∞mean␈α∞the␈α∞same␈α∞thing␈α∞for␈α∞all␈α∞types␈α∞of
␈↓ α←␈↓data␈α
structures),␈α
this␈α
is␈α
not␈α
critical␈α
to␈α
the␈α
formalism.␈α
 Slotnames␈α
could␈α
easily
␈↓ α←␈↓be␈α∩made␈α∩local␈α∪to␈α∩a␈α∩given␈α∩schema,␈α∪and␈α∩the␈α∩schema␈α∩name␈α∪would␈α∩become
␈↓ α←␈↓another␈α∞index␈α∞in␈α∞the␈α∞knowledge␈α∞organization␈α∞framework.␈α∞ Thus,␈α∂instead␈α∞of
␈↓ α←␈↓indexing␈αthe␈αknowledge␈αin␈αthe␈αslotexperts␈αby␈αslotname␈αand␈αadvice,␈αwe␈αwould
␈↓ α←␈↓index␈α
by␈αschema␈α
name,␈αslotname,␈α
and␈αadvice.␈α
The␈αpower␈α
and␈α
limitations␈αof
␈↓ α←␈↓the framework would remain unchanged.
␈↓ α←␈↓␈↓6-8␈↓ ∧SKNOWLEDGE ABOUT REPRESENTATIONS:  ORGANIZATION    153␈↓

␈↓"β␈↓ α←␈↓everything␈α∀in␈α∀the␈α∀system␈α∀should␈α∪be␈α∀an␈α∀instance␈α∀of␈α∀some␈α∀schema.␈α∪ One
␈↓ α←␈↓implication␈α∞of␈α∂this␈α∞was␈α∞that␈α∂both␈α∞the␈α∞schemata␈α∂and␈α∞their␈α∂components␈α∞were
␈↓ α←␈↓themselves␈α
considered␈αextended␈α
data␈αtypes.␈α
 Evidence␈αof␈α
this␈αcan␈α
be␈α
seen␈αin
␈↓ α←␈↓the␈αslotnames.␈α
 There␈αis␈α
a␈α␈↓	SLOTNAME-SCHEMA␈↓␈α
that␈αdescribes␈α
the␈αstructure␈αof␈α
a
␈↓ α←␈↓slotname and makes it possible to acquire new slotnames interactively.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∩of␈α∩the␈α∩consequences␈α∩of␈α∩this␈α∩approach␈α∩is␈α∩a␈α∩circularity␈α∩in␈α∩the
␈↓ α←␈↓definitions␈α∞of␈α∞the␈α∂data␈α∞types.␈α∞ For␈α∞instance,␈α∂␈↓	DESCR␈↓iption␈α∞is␈α∞a␈α∂slotname␈α∞and,
␈↓ α←␈↓hence,␈α∂an␈α∂instance␈α∂of␈α∂the␈α⊂␈↓	SLOTNAME-SCHEMA␈↓.␈α∂ But␈α∂part␈α∂of␈α∂the␈α⊂structure␈α∂of
␈↓ α←␈↓every␈α⊂slotname␈α⊃is␈α⊂a␈α⊃␈↓	DESCR␈↓iption␈α⊂specifying␈α⊃what␈α⊂that␈α⊃slotname␈α⊂represents.
␈↓ α←␈↓Hence,␈αthere␈αis␈αa␈α␈↓	DESCR␈↓iption␈αof␈α␈↓	DESCR␈↓iption.␈α Similarly,␈αin␈αacquiring␈αa␈αnew
␈↓ α←␈↓slotname,␈α
the␈αsystem␈α
requests␈αa␈α
prompt␈α
for␈αit,␈α
using␈αthe␈α
prompt␈α
for␈α␈↓	PROMPT␈↓:
␈↓ α←␈↓␈↓↓Please␈αgive␈αme␈αa␈α
short␈αphrase␈αwhich␈αcan␈α
be␈αused␈αto␈αask␈α
for␈αthe␈αcontents␈αof␈α
this
␈↓ α←␈↓↓slot␈↓.␈α
 This␈αcircularity␈α
is␈α
a␈αresult␈α
of␈α
the␈αsystematic␈α
application␈α
of␈αthe␈α
use␈αof␈α
the
␈↓ α←␈↓extended␈α∞data␈α
types␈α∞and␈α∞makes␈α
possible␈α∞the␈α
sort␈α∞of␈α∞bootstrapping␈α
behavior
␈↓ α←␈↓demonstrated later in this chapter.
␈↓ α←␈↓␈↓154    KNOWLEDGE ACQUISITION II␈↓ 
#6-9␈↓

␈↓"β␈↓ α←␈↓␈↓α6-9    KNOWLEDGE ABOUT REPRESENTATIONS:  USE␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Section␈α6-8␈α
described␈αthe␈α
organization␈αand␈α
content␈αof␈α
the␈αknowledge
␈↓ α←␈↓about␈α⊂representations␈α⊂embodied␈α⊂in␈α⊂the␈α⊂schemata␈α⊂and␈α⊃associated␈α⊂structures.
␈↓ α←␈↓This␈αsection␈αdescribes␈αhow␈αthat␈αinformation␈αis␈αused;␈αin␈αparticular,␈αthe␈αway␈αit
␈↓ α←␈↓enables␈α∪the␈α∩expert␈α∪to␈α∩teach␈α∪the␈α∩system␈α∪about␈α∩new␈α∪conceptual␈α∩primitives.
␈↓ α←␈↓Other uses (e.g., for information storage and retrieval) are also described.

␈↓"β␈↓ α←␈↓␈↓α6-9-1    Schema function:  Acquisition of new instances␈↓
␈↓"β␈↓ α←␈↓␈↓ β?We␈α
begin␈α∞at␈α
the␈α
point␈α∞where␈α
some␈α
schema␈α∞in␈α
the␈α
network␈α∞has␈α
been
␈↓ α←␈↓selected␈α∩as␈α∩a␈α∩starting␈α∩point␈α∩(Section␈α∩6-9-2␈α∩discusses␈α∩how␈α∩this␈α∩decision␈α⊃is
␈↓ α←␈↓made).␈α⊂ Since␈α⊃information␈α⊂is␈α⊂distributed␈α⊃through␈α⊂the␈α⊂schema␈α⊃network,␈α⊂the
␈↓ α←␈↓first␈α∂step␈α∂is␈α⊂to␈α∂get␈α∂to␈α⊂the␈α∂root,␈α∂keeping␈α∂track␈α⊂of␈α∂the␈α∂path␈α⊂while␈α∂ascending.
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓␈α
``climbs''␈α
up␈α
the␈α
␈↓	FATHER␈↓␈α
links,␈α
marking␈α
each␈α
schema␈α
along␈αthe␈α
way.␈↓
11␈↓
␈↓ α←␈↓The␈α
system␈αeventually␈α
arrives␈αat␈α
the␈αroot,␈α
with␈αall␈α
or␈αsome␈α
part␈αof␈α
the␈αpath
␈↓ α←␈↓marked␈α∂back␈α∂down␈α∞to␈α∂a␈α∂terminal␈α∂schema.␈α∞ (Parts␈α∂may␈α∂be␈α∂unmarked␈α∞either
␈↓ α←␈↓because␈α⊂it␈α⊂jumped␈α⊂over␈α⊂non-unique␈α⊂parents␈α⊂or␈α⊂because␈α⊂the␈α⊂starting␈α⊂point
␈↓ α←␈↓chosen␈αwas␈αnot␈αa␈αterminal␈αof␈αthe␈αnetwork.␈α The␈αlatter␈αcase␈αwould␈αarise␈αif,␈αfor
␈↓ α←␈↓instance,␈α␈↓¬TEIRESIAS␈↓␈αknew␈αonly␈αthat␈αthe␈αexpert␈αwanted␈αto␈αcreate␈αa␈αnew␈α
kind␈αof
␈↓ α←␈↓value but was not able to discover which type.)
␈↓"β␈↓ α←␈↓␈↓ β?The␈αnext␈αstep␈α
is␈αto␈αdescend␈α
back␈αdown␈αthe␈α
network␈αalong␈αthe␈α
marked
␈↓ α←␈↓path,␈α∞using␈α∞each␈α∞schema␈α∂along␈α∞the␈α∞way␈α∞as␈α∂a␈α∞further␈α∞set␈α∞of␈α∂instructions␈α∞for
␈↓ α←␈↓acquiring␈αthe␈αnew␈αinstance.␈α If␈αthe␈αprocess␈αencounters␈αa␈αpart␈αof␈αthe␈αpath␈α
that
␈↓ α←␈↓is␈αnot␈αmarked,␈αthe␈αexpert's␈αhelp␈αis␈αrequested.␈α This␈αis␈αdone␈αby␈αdisplaying␈αthe
␈↓ α←␈↓English␈α∂phrase␈α∂(the␈α⊂␈↓	STRAN␈↓)␈α∂associated␈α∂with␈α∂each␈α⊂of␈α∂the␈α∂␈↓	OFFSPRING␈↓␈α⊂of␈α∂the
␈↓ α←␈↓current␈αschema␈αand␈αasking␈αthe␈αexpert␈αto␈αchoose␈αthe␈αone␈αwhich␈αbest␈αdescribes
␈↓ α←␈↓the item being constructed.
␈↓"β␈↓ α←␈↓␈↓ β?At␈α
each␈αnode␈α
in␈αthe␈α
network␈α
the␈αacquisition␈α
process␈αis␈α
directed␈α
by␈αa
␈↓ α←␈↓simple␈α⊂``schema␈α⊃interpreter''␈α⊂whose␈α⊂control␈α⊃structure␈α⊂consists␈α⊂of␈α⊃three␈α⊂basic
␈↓ α←␈↓operations:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?Use␈α
the␈α
structure␈α
description␈α
part␈αof␈α
the␈α
schema␈α
to␈α
guide␈αthe
␈↓ α←␈↓␈↓ β?addition of new components to the instance,

␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?attend␈αto␈αany␈αupdating␈αaccording␈αto␈αthe␈α
information␈αspecified
␈↓ α←␈↓␈↓ β?in the ␈↓	RELATIONS␈↓, and



␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[11]␈αIf␈αit␈αencounters␈αa␈αschema␈α
that␈αhas␈αmultiple␈αparents,␈αit␈αjumps␈α
directly␈αto
␈↓ α←␈↓the␈α∞network␈α∞root.␈α∞ This␈α∞is␈α
a␈α∞sub-optimal␈α∞solution;␈α∞a␈α∞better␈α∞approach␈α
would
␈↓ α←␈↓have␈αa␈α
more␈αsophisticated␈α
treatment␈αof␈αthe␈α
network.␈αIt␈α
might,␈αfor␈αinstance,␈α
be
␈↓ α←␈↓able␈α∩to␈α∪recognize␈α∩the␈α∩situation␈α∪in␈α∩which␈α∩all␈α∪the␈α∩parents␈α∩had␈α∪a␈α∩common
␈↓ α←␈↓``grandparent''␈α∞and␈α∞thus␈α∞jump␈α∞only␈α
two␈α∞levels␈α∞(over␈α∞the␈α∞ambiguous␈α
section),
␈↓ α←␈↓rather than straight to the root.
␈↓ α←␈↓␈↓6-9␈↓ ¬XKNOWLEDGE ABOUT REPRESENTATIONS:  USE    155␈↓

␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?add the new item to the schema's list of instances.␈↓
12␈↓


␈↓"β␈↓ α←␈↓␈↓αAdding to the structure of the new concept␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
process␈αof␈α
adding␈αnew␈α
components␈αto␈α
the␈αnew␈α
instance␈αinvolves
␈↓ α←␈↓filling␈α↔in␈α⊗slots,␈α↔as␈α↔guided␈α⊗by␈α↔the␈α↔information␈α⊗provided␈α↔in␈α↔the␈α⊗␈↓↓blank␈↓.
␈↓ α←␈↓Computationally,␈α∃the␈α∃process␈α∃involves␈α∃sending␈α∃the␈α∃␈↓↓blank␈↓␈α∃and␈α⊗␈↓↓advice␈↓␈α∃as
␈↓ α←␈↓arguments to the appropriate slot expert:␈↓
13␈↓

␈↓"β␈↓ α←␈↓	␈↓ βW(APPLY* (GETEXPERT <slotname>) <blank> <advice>)

␈↓ α←␈↓The␈α⊃segment␈α⊃of␈α⊃code␈α⊃in␈α⊂the␈α⊃␈↓	SLOTEXPERT␈↓␈α⊃associated␈α⊃with␈α⊃the␈α⊃␈↓↓advice␈↓␈α⊂then
␈↓ α←␈↓determines␈α∂how␈α∂to␈α∂go␈α∂about␈α∂filling␈α∂in␈α∂the␈α∂blank.␈α∂ For␈α∂example,␈α∂when␈α∞that
␈↓ α←␈↓␈↓↓advice␈↓␈α
is␈α␈↓	ASKIT␈↓,␈α
the␈αexpert␈α
is␈αconsulted.␈α
 As␈α
we␈αhave␈α
seen,␈αthis␈α
appears␈αto␈α
the
␈↓ α←␈↓expert␈α⊃as␈α⊃a␈α⊃process␈α⊃of␈α⊃supplying␈α⊃information␈α⊃in␈α⊃a␈α⊃form␈α⊃specified␈α∩by␈α⊃the
␈↓ α←␈↓system: ␈α␈↓¬TEIRESIAS␈↓␈α
first␈αprints␈αa␈α
``translated''␈αversion␈αof␈α
the␈α␈↓↓blank␈↓␈αto␈α
guide␈αthe
␈↓ α←␈↓expert, then uses the same ␈↓↓blank␈↓ to parse his response.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α
approach␈α
makes␈αpossible␈α
a␈α
particularly␈αsimple␈α
form␈α
of␈α``schema
␈↓ α←␈↓interpreter.'' ␈α∀For␈α∀instance,␈α∀the␈α∀part␈α∪of␈α∀the␈α∀interpreter␈α∀that␈α∀handles␈α∪this
␈↓ α←␈↓addition␈α⊂of␈α⊂new␈α⊂substructure␈α⊂is␈α⊂just␈α∂the␈α⊂single␈α⊂line␈α⊂of␈α⊂code␈α⊂shown␈α∂above.
␈↓ α←␈↓The␈αtask␈αof␈αfilling␈αin␈αthe␈αblank␈α
is␈αthus␈αhanded␈αoff␈αto␈αthe␈αappropriate␈α
␈↓↓expert␈↓.
␈↓ α←␈↓The␈α⊂␈↓↓expert␈↓,␈α⊃in␈α⊂turn,␈α⊂hands␈α⊃it␈α⊂to␈α⊃the␈α⊂segment␈α⊂of␈α⊃code␈α⊂associated␈α⊃with␈α⊂the
␈↓ α←␈↓indicated␈α
piece␈α
of␈α
␈↓↓advice␈↓.␈α
 That␈α
code␈α
may,␈α
in␈α
turn,␈α
request␈α
each␈α
part␈α∞of␈α
the
␈↓ α←␈↓␈↓↓blank␈↓␈αto␈αsupply␈αa␈α``translation''␈αof␈αitself␈αfor␈αdisplay␈αto␈αthe␈αuser.␈α Thus,␈αrather
␈↓ α←␈↓than␈αtrying␈αto␈αwrite␈α
a␈αclever␈αinterpreter␈αthat␈α
had␈αa␈αlot␈αof␈α
information␈αabout
␈↓ α←␈↓each␈α∞representation,␈α
we␈α∞have␈α
instead␈α∞written␈α
a␈α∞simple␈α
interpreter␈α∞and␈α
allow
␈↓ α←␈↓the representations themselves to supply the information.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α⊂are␈α⊂also␈α⊂several␈α⊂human␈α⊂engineering␈α⊂features␈α⊂available␈α∂when
␈↓ α←␈↓the␈αadvice␈α
indicates␈αthat␈αthe␈α
information␈αto␈αfill␈α
the␈αslot␈αshould␈α
be␈αrequested
␈↓ α←␈↓from␈α
the␈α
expert.␈α∞ We␈α
have␈α
seen␈α∞the␈α
use␈α
of␈α
the␈α∞␈↓↓blank␈↓␈α
in␈α
guiding␈α∞the␈α
expert
␈↓ α←␈↓and␈α⊂in␈α⊂parsing␈α⊂his␈α⊂answer.␈α⊂ There␈α⊂is␈α⊂also␈α⊂the␈α⊂ability␈α⊂to␈α⊂display␈α⊂a␈α∂sample
␈↓ α←␈↓answer␈α(in␈αresponse␈αto␈αa␈α``?''␈α)␈αor␈αall␈αlegal␈αanswers␈α(in␈αresponse␈αto␈αa␈α``??'').␈α All
␈↓ α←␈↓of these help to make the interaction relatively painless.

␈↓"β␈↓ α←␈↓␈↓αAttending to data structure interrelations␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∞next␈α∞step--dealing␈α∂with␈α∞necessary␈α∞updates␈α∞to␈α∂other␈α∞structures--

␈↓ α←␈↓_______________________________

␈↓"β␈↓ α←␈↓[12]␈α∞For␈α∞the␈α∞sake␈α∞of␈α∞efficiency,␈α∂only␈α∞schemata␈α∞at␈α∞the␈α∞leaves␈α∞of␈α∂the␈α∞network
␈↓ α←␈↓keep␈αtrack␈αof␈α
instances.␈α Each␈αnew␈α
item␈αcarries␈αa␈α
record␈αof␈αits␈α
path␈αthrough
␈↓ α←␈↓the␈α⊃network␈α⊃(in␈α∩its␈α⊃␈↓	INSTOF␈↓␈α⊃property);␈α∩this␈α⊃allows␈α⊃disambiguation␈α∩when␈α⊃a
␈↓ α←␈↓schema has more than one parent in the network.

␈↓"β␈↓ α←␈↓[13] ␈↓	APPLY*␈↓ applies its first argument to its remaining arguments.  Thus,
␈↓"β␈↓ α←␈↓	␈↓ βw(APPLY* (QUOTE CONS) (QUOTE A) NIL) = (A)  .
␈↓ α←␈↓␈↓156    KNOWLEDGE ACQUISITION II␈↓ 
#6-9␈↓

␈↓"β␈↓ α←␈↓relies␈αon␈αthe␈αinformation␈αspecified␈αin␈αthe␈α␈↓	RELATIONS␈↓␈αslot.␈α The␈αbasic␈αidea␈αis
␈↓ α←␈↓to␈α∪consider␈α∪this␈α∪information␈α∪as␈α∀a␈α∪list␈α∪of␈α∪potential␈α∪updating␈α∪tasks␈α∀to␈α∪be
␈↓ α←␈↓performed whenever a new instance of the schema is acquired.
␈↓"β␈↓ α←␈↓␈↓ β?Maintaining␈α
existing␈αinterdependencies␈α
of␈αdata␈α
structures␈αin␈α
the␈αface
␈↓ α←␈↓of additions to the system requires three kinds of information:

␈↓"β␈↓ α←␈↓␈↓ ββ(1)␈↓ β?What␈α
other␈α
structures␈α
might␈α
need␈α
to␈α
be␈α
updated␈α
in␈α
response␈α
to
␈↓ α←␈↓␈↓ β?the new addition?
␈↓"β␈↓ α←␈↓␈↓ ββ(2)␈↓ β?If␈α→those␈α→other␈α→structures␈α_are␈α→not␈α→all␈α→independent,␈α_what
␈↓ α←␈↓␈↓ β?interrelationships exist between them?
␈↓"β␈↓ α←␈↓␈↓ ββ(3)␈↓ β?What effect should the new addition have on each structure?

␈↓"β␈↓ α←␈↓␈↓ β?As␈αan␈α
example,␈αconsider␈αthe␈α
acquisition␈αof␈αthe␈α
new␈αculture␈αsite␈α
shown
␈↓ α←␈↓earlier.␈α
 The␈α
first␈α
updating␈α
task␈α
encountered␈α
is␈α
the␈α
decision␈α
whether␈α
to␈αadd
␈↓ α←␈↓the␈αnew␈αsite␈αto␈αthe␈αcollection␈αof␈αsterile,␈αnonsterile,␈αor␈αindeterminant␈αsites.␈α We
␈↓ α←␈↓describe␈αthis␈αby␈αsaying␈αthat␈αthe␈αdata␈αtype␈α␈↓	SITE␈↓␈αis␈αthe␈α``trigger''␈αfor␈αan␈αaction
␈↓ α←␈↓that␈αmay␈αneed␈αto␈αbe␈αperformed␈αon␈αone␈αor␈αmore␈α``targets.''␈αIn␈αthese␈α
terms,␈αthe
␈↓ α←␈↓targets␈α
are␈α
the␈α
answers␈αto␈α
question␈α
1␈α
above␈α(other␈α
structures␈α
that␈α
may␈αneed␈α
to
␈↓ α←␈↓be␈α∂updated),␈α∂and␈α∂the␈α∂fact␈α∂that␈α∞the␈α∂categories␈α∂are␈α∂mutually␈α∂exclusive␈α∂is␈α∞the
␈↓ α←␈↓answer␈α
to␈α
question␈α2␈α
(the␈α
constraints␈α
on␈αthe␈α
effects).␈α
 Information␈α
needed␈αto
␈↓ α←␈↓answer␈α∂question␈α⊂3␈α∂(the␈α⊂effect␈α∂on␈α∂each␈α⊂target)␈α∂may␈α⊂come␈α∂from␈α⊂two␈α∂sources.
␈↓ α←␈↓First,␈α⊂the␈α⊃data␈α⊂type␈α⊃and␈α⊂organization␈α⊂of␈α⊃the␈α⊂target␈α⊃is␈α⊂always␈α⊃relevant.␈α⊂ A
␈↓ α←␈↓partially␈α
ordered␈αlist,␈α
for␈αexample,␈α
will␈αbe␈α
updated␈αone␈α
way,␈αwhile␈α
a␈α
set␈αwill
␈↓ α←␈↓be␈α∞updated␈α∞in␈α∞another.␈α∞ Second,␈α∞the␈α∞trigger␈α∞may␈α∞or␈α∞may␈α∞not␈α∞carry␈α
relevant
␈↓ α←␈↓information.␈α
 In␈α
the␈α∞example␈α
above,␈α
it␈α∞does␈α
not.␈α
 Adding␈α∞a␈α
new␈α
site␈α∞to␈α
any
␈↓ α←␈↓one␈α
of␈α
the␈α
three␈α
categories␈α
requires␈α
no␈α
information␈α
about␈α
the␈α
site␈α∞itself;␈α
the
␈↓ α←␈↓system␈α∞need␈α∞only␈α
know␈α∞to␈α∞which␈α
target␈α∞it␈α∞should␈α
be␈α∞added.␈α∞ The␈α
approach
␈↓ α←␈↓used␈α∂here␈α∂is␈α∂particularly␈α∂well␈α∂suited␈α∂to␈α∂this␈α∂situation␈α∂(in␈α∂which␈α∂the␈α∂trigger
␈↓ α←␈↓does␈α∂not␈α⊂determine␈α∂the␈α∂effect␈α⊂on␈α∂the␈α⊂target)␈α∂and␈α∂takes␈α⊂advantage␈α∂of␈α⊂it␈α∂by
␈↓ α←␈↓minimizing␈α⊂the␈α⊂distribution␈α⊂of␈α⊂the␈α⊂required␈α⊂knowledge␈α⊂(this␈α⊂point␈α⊃will␈α⊂be
␈↓ α←␈↓clarified below).
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∀``language''␈α∀of␈α∀the␈α∀␈↓	RELATIONS␈↓␈α∀is␈α∀a␈α∀syntax␈α∀of␈α∀data␈α∪structure
␈↓ α←␈↓interrelationships␈αand␈αprovides␈αa␈αway␈αof␈αexpressing␈αthe␈αanswers␈αto␈αquestions
␈↓ α←␈↓1␈α⊂and␈α⊃2.␈α⊂ For␈α⊃the␈α⊂current␈α⊃example,␈α⊂part␈α⊃of␈α⊂the␈α⊃␈↓	RELATIONS␈↓␈α⊂of␈α⊃the␈α⊂␈↓	SITE-
␈↓ α←␈↓	SCHEMA␈↓ is

␈↓"β␈↓ α←␈↓	␈↓ βg(XOR* STERILESITES NONSTERILESITES OTHERSITES)

␈↓ α←␈↓which␈α
indicates␈α
which␈α
structures␈α
are␈α
potentially␈α
affected␈α
and␈α
the␈αconstraint␈α
of
␈↓ α←␈↓mutual␈α∞exclusion.␈α∞ The␈α∞information␈α∞for␈α∞question␈α∞3␈α∞is␈α∞supplied␈α∞by␈α
updating
␈↓ α←␈↓functions (described below), which are included in some of the schemata.
␈↓"β␈↓ α←␈↓␈↓ β?One␈αexample␈αof␈αhow␈α
the␈αupdating␈αprocess␈αworks␈α
will␈αmake␈αall␈αof␈α
this
␈↓ α←␈↓clearer␈α
and␈αillustrate␈α
the␈αadvantages␈α
it␈α
presents.␈αThe␈α
first␈αstep␈α
is␈αto␈α
determine
␈↓ α←␈↓which␈α⊃structures␈α⊃should␈α⊃actually␈α⊃be␈α⊃updated.␈α⊃ If␈α⊃the␈α⊃␈↓	<switch>␈↓␈α⊃is␈α⊃␈↓	OR*␈↓␈α⊃or
␈↓ α←␈↓␈↓6-9␈↓ ¬XKNOWLEDGE ABOUT REPRESENTATIONS:  USE    157␈↓

␈↓"β␈↓ α←␈↓␈↓	XOR*␈↓,␈α∃the␈α∃expert's␈α∃help␈α⊗is␈α∃requested;␈↓
14␈↓␈α∃otherwise␈α∃(␈↓	AND*␈↓,␈α∃or␈α⊗a␈α∃recursive
␈↓ α←␈↓definition),␈α∞the␈α∞system␈α∞itself␈α
can␈α∞make␈α∞the␈α∞decision.␈α
 In␈α∞this␈α∞case␈α∞the␈α
system
␈↓ α←␈↓displays␈αthe␈αthree␈αchoices␈α(␈↓↓sterile,␈αnonsterile,␈α␈↓and␈↓↓␈αindeterminant␈↓)␈αand␈αasks␈αthe
␈↓ α←␈↓expert to select one.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α↔rest␈α_of␈α↔the␈α↔process␈α_can␈α↔best␈α↔be␈α_viewed␈α↔by␈α_adopting␈α↔the
␈↓ α←␈↓perspective␈α⊂of␈α⊂much␈α⊂of␈α⊂the␈α⊂work␈α⊂on␈α⊂``actors''␈α⊂[Hewitt75]␈α⊂and␈α⊃the␈α⊂␈↓¬SMALLTALK␈↓
␈↓ α←␈↓language␈α↔[Learning76],␈α↔in␈α↔which␈α↔data␈α↔structures␈α↔are␈α_considered␈α↔active
␈↓ α←␈↓elements␈α∞that␈α
exchange␈α∞messages.␈α
 In␈α∞these␈α∞terms,␈α
the␈α∞next␈α
step␈α∞is␈α∞to␈α
``send''
␈↓ α←␈↓the␈α∩new␈α∪culture␈α∩site␈α∪to␈α∩the␈α∪target␈α∩selected␈α∪(␈↓	OTHERSITES␈↓),␈α∩along␈α∪with␈α∩the
␈↓ α←␈↓command␈α∂to␈α∞the␈α∂target␈α∞to␈α∂``Add␈α∂this␈α∞to␈α∂yourself.'' ␈α∞The␈α∂target␈α∂``knows''␈α∞that
␈↓ α←␈↓knowledge␈α∂about␈α∂its␈α∂structure␈α⊂is␈α∂stored␈α∂with␈α∂the␈α⊂schema␈α∂of␈α∂which␈α∂it␈α⊂is␈α∂an
␈↓ α←␈↓instance,␈α∞so␈α∞it␈α∞finds␈α∞a␈α∞way␈α∞to␈α∞pass␈α∞the␈α∞buck: ␈α∞It␈α∞examines␈α∞itself␈α∞to␈α∞find␈α∞out
␈↓ α←␈↓which␈αschema␈αit␈αis␈αan␈αinstance␈αof␈α(i.e.,␈αit␈αexamines␈αthe␈αcontents␈αof␈αits␈α␈↓	INSTOF␈↓
␈↓ α←␈↓slot).␈α⊂ Determining␈α⊃that␈α⊂it␈α⊂is␈α⊃an␈α⊂instance␈α⊂of␈α⊃the␈α⊂schema␈α⊃for␈α⊂alphabetically
␈↓ α←␈↓ordered␈α⊃linear␈α⊂lists␈α⊃(the␈α⊃␈↓	AOLL-SCHEMA␈↓),␈α⊂it␈α⊃sends␈α⊃a␈α⊂request␈α⊃to␈α⊃this␈α⊂schema,
␈↓ α←␈↓asking the schema to take care of the ``add this'' message.
␈↓"β␈↓ α←␈↓␈↓ β?Recall␈α⊂that␈α∂the␈α⊂schema␈α∂is␈α⊂a␈α∂device␈α⊂for␈α∂organizing␈α⊂a␈α∂wide␈α⊂range␈α∂of
␈↓ α←␈↓information␈αabout␈αrepresentations.␈α Part␈α
of␈αthat␈αinformation␈αindicates␈αhow␈α
to
␈↓ α←␈↓augment␈α⊃existing␈α⊃data␈α⊃structures.␈α∩ The␈α⊃␈↓	AOLL-SCHEMA␈↓␈α⊃(like␈α⊃others)␈α∩has␈α⊃an
␈↓ α←␈↓``updating␈α→function''␈α→capable␈α→of␈α→adding␈α→new␈α→elements␈α→to␈α→its␈α_instances
␈↓ α←␈↓(alphabetically␈α
ordered␈α
linear␈α
lists)␈αwithout␈α
violating␈α
their␈α
established␈αorder.
␈↓ α←␈↓Thus,␈α
in␈α
response␈α
to␈α
the␈α
request␈α
from␈α
␈↓	OTHERSITES␈↓,␈α
the␈α
␈↓	AOLL-SCHEMA␈↓␈α
invokes
␈↓ α←␈↓its␈α
updating␈α
function␈α
on␈α
the␈α
new␈α
culture␈α
site␈α
and␈α
the␈α
list␈α␈↓	OTHERSITES␈↓,␈α
adding
␈↓ α←␈↓the new element to the list in the proper place.
␈↓"β␈↓ α←␈↓␈↓ β?To review:

␈↓"β␈↓ α←␈↓␈↓ β'␈↓	SITE-SCHEMA␈↓␈α
asks␈α
the␈α
expert␈α
if␈α
the␈α
new␈α
site␈α
is␈α∞␈↓↓sterile␈↓,␈α
␈↓↓nonsterile␈↓,
␈↓ α←␈↓␈↓ β'or ␈↓↓indeterminant␈↓.
␈↓"β␈↓ α←␈↓␈↓ β'The expert indicates ␈↓↓indeterminant␈↓.
␈↓"β␈↓ α←␈↓␈↓ β'␈↓	SITE-SCHEMA␈↓␈αsends␈αthe␈αnew␈α
site␈αto␈α␈↓	OTHERSITES␈↓,␈αwith␈αthe␈α
message
␈↓ α←␈↓␈↓ β'``Add this to yourself.''
␈↓"β␈↓ α←␈↓␈↓ β'␈↓	OTHERSITES␈↓␈α∪examines␈α∀itself,␈α∪finds␈α∪it␈α∀is␈α∪an␈α∪instance␈α∀of␈α∪␈↓	AOLL-
␈↓ α←␈↓	␈↓ β'SCHEMA␈↓,␈α∂and␈α∞sends␈α∂a␈α∞message␈α∂to␈α∞␈↓	AOLL-SCHEMA␈↓␈α∂saying␈α∂``Add␈α∞this
␈↓ α←␈↓␈↓ β'new site to me.''
␈↓"β␈↓ α←␈↓␈↓ β'The␈α∩updating␈α∩function␈α∩associated␈α∩with␈α∩␈↓	AOLL-SCHEMA␈↓␈α∪adds␈α∩the
␈↓ α←␈↓␈↓ β'new site to ␈↓	OTHERSITES␈↓.

␈↓"β␈↓ α←␈↓␈↓ β?There␈α∂are␈α∂several␈α∂advantages␈α⊂to␈α∂the␈α∂distribution␈α∂of␈α⊂knowledge␈α∂this
␈↓ α←␈↓technique␈α
employs.␈α
 To␈α
make␈α
them␈αclear,␈α
consider␈α
the␈α
generalized␈α
view␈αof␈α
the
␈↓ α←␈↓process␈α
shown␈αin␈α
Fig.␈α6-7.␈α
 Shown␈α
there␈αis␈α
one␈αtrigger␈α
(␈↓	SITE␈↓),␈α
three␈αtargets,

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[14]␈α⊂Recall␈α∂that␈α⊂all␈α⊂data␈α∂structures␈α⊂in␈α⊂the␈α∂knowledge␈α⊂base␈α⊂have␈α∂associated
␈↓ α←␈↓with␈αthem␈αa␈αdescriptive␈αEnglish␈αphrase␈α(the␈α␈↓	DESCR␈↓␈αpart)␈αsupplied␈αduring␈αthe
␈↓ α←␈↓acquisition␈αprocess.␈α It␈αis␈αthis␈αdescription␈αthat␈αallows␈α␈↓¬TEIRESIAS␈↓␈αto␈α``talk''␈αabout
␈↓ α←␈↓various data structures.
␈↓ α←␈↓␈↓158    KNOWLEDGE ACQUISITION II␈↓ 
#6-9␈↓

␈↓"β␈↓ α←␈↓and␈α
a␈α
structure␈α
called␈α
a␈α
``traffic␈α
director''␈α
(which␈α
is␈α
a␈α
generalized␈α∞version␈α
of
␈↓ α←␈↓the␈α∪␈↓	RELATIONS␈↓).␈α∩ In␈α∪this␈α∪view,␈α∩each␈α∪schema␈α∪would␈α∩have␈α∪its␈α∪own␈α∩traffic
␈↓ α←␈↓director␈α∩that␈α∩tells␈α∩it␈α∩what␈α∩to␈α∩do␈α∩with␈α∩new␈α∩instances.␈α∩ The␈α∩basic␈α∪issue␈α∩is
␈↓ α←␈↓organization␈αof␈αknowledge␈αand,␈αin␈αparticular,␈αhow␈αthat␈αknowledge␈αshould␈αbe
␈↓ α←␈↓distributed between the updating functions and the traffic director.

␈↓"␈↓ α←␈↓∧  ⊂αααααααααααααα⊃     ⊂ααααααααααααααααα⊃     ⊂αααααααααααα⊃
␈↓"␈↓ α←␈↓∧  ~   updating   ~     ~    updating     ~     ~  updating  ~
␈↓"␈↓ α←␈↓∧  ~  function-1  ~     ~   function-2    ~     ~ function-3 ~
␈↓"␈↓ α←␈↓∧  ε α α α α α α αλ     ε α α α α α α α α λ     ε α α α α α αλ
␈↓"␈↓ α←␈↓∧  ~   target-1   ~     ~    target-2     ~     ~  target-3  ~
␈↓"␈↓ α←␈↓∧  ~              ~     ~                 ~     ~            ~
␈↓"␈↓ α←␈↓∧  ~ STERILESITES ~     ~ NONSTERILESITES ~     ~ OTHERSITES ~
␈↓"␈↓ α←␈↓∧  %αααααααααααααα$     %ααααααααααααααααα$     %αααααααααααα$




␈↓"␈↓ α←␈↓∧                       ⊂αααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧                       ~ traffic director ~
␈↓"␈↓ α←␈↓∧                       ε α α α α α α α α αλ
␈↓"␈↓ α←␈↓∧                       ~   SITE-SCHEMA    ~
␈↓"␈↓ α←␈↓∧                       %αααααααααααααααααα$

␈↓"β␈↓ α←␈↓α␈↓ β≠Fig.␈α~6-7.    Generalized␈α→view␈α~of␈α→attending␈α~to␈α~data␈α→structure
␈↓ α←␈↓α␈↓ β≠interrelations.    

␈↓"β␈↓ α←␈↓␈↓ β?In␈α⊃the␈α∩current␈α⊃example,␈α⊃a␈α∩new␈α⊃culture␈α⊃site␈α∩is␈α⊃``sent''␈α⊃to␈α∩the␈α⊃traffic
␈↓ α←␈↓director␈α~for␈α~instructions.␈α~ It␈α→might␈α~receive␈α~three␈α~kinds␈α~of␈α→directions,
␈↓ α←␈↓depending␈α∂on␈α∂how␈α⊂much␈α∂information␈α∂is␈α⊂stored␈α∂there.␈α∂ The␈α⊂traffic␈α∂director
␈↓ α←␈↓might know:

␈↓"β␈↓ α←␈↓DESIGN A:␈↓ ∧∪the names of the targets.
␈↓"β␈↓ α←␈↓DESIGN B:␈↓ ∧∪the names of the targets and the constraints among them.
␈↓"β␈↓ α←␈↓DESIGN C:␈↓ ∧∪the␈α
names␈α
of␈α
the␈α
targets,␈α
the␈α
constraints␈α
among␈α
them,␈α
and␈α
the
␈↓ α←␈↓␈↓ ∧∪structure of each target.

␈↓"β␈↓ α←␈↓␈↓ β?Design␈α_C␈α_organizes␈α_the␈α_updating␈α_process␈α_around␈α→each␈α_trigger,
␈↓ α←␈↓corresponding␈αclosely␈αto␈αthe␈αstandard␈αdemon-like␈αapproach.␈α In␈αthis␈αcase␈αthe
␈↓ α←␈↓traffic␈α
director␈α
can␈α``tell''␈α
the␈α
new␈αsite␈α
exactly␈α
which␈α
target(s)␈αto␈α
go␈α
to␈αand␈α
how
␈↓ α←␈↓to␈α``add␈α
itself''␈αto␈αeach.␈α
This␈αwould␈αmean␈α
most␈αof␈αthe␈α
knowledge␈αis␈α
stored␈αin
␈↓ α←␈↓the␈α∀traffic␈α∀director,␈α∀which␈α∀has␈α∀to␈α∀know␈α∀both␈α∀the␈α∀organization␈α∀and␈α∀the
␈↓ α←␈↓structure of all current targets.
␈↓"β␈↓ α←␈↓␈↓ β?Design␈α∂A␈α∞organizes␈α∂the␈α∞process␈α∂around␈α∞the␈α∂target.␈α∞ Here␈α∂the␈α∞traffic
␈↓ α←␈↓director␈αcan␈α
only␈αsay␈α``Here's␈α
all␈αthe␈α
places␈αyou␈αmight␈α
(or␈αmight␈α
not)␈αbelong,
␈↓ α←␈↓try␈α⊃them␈α⊃all␈α⊃and␈α⊃ask␈α⊂when␈α⊃you␈α⊃get␈α⊃there.'' ␈α⊃In␈α⊂this␈α⊃case␈α⊃the␈α⊃bulk␈α⊃of␈α⊂the
␈↓ α←␈↓knowledge␈α∞is␈α∞stored␈α∞with␈α∂the␈α∞updating␈α∞functions: ␈α∞They␈α∞will␈α∂be␈α∞responsible
␈↓ α←␈↓for␈α⊃adding␈α∩the␈α⊃new␈α∩site␈α⊃to␈α⊃the␈α∩target␈α⊃and␈α∩for␈α⊃maintaining␈α∩the␈α⊃necessary
␈↓ α←␈↓interrelations between the targets.
␈↓"β␈↓ α←␈↓␈↓6-9␈↓ ¬XKNOWLEDGE ABOUT REPRESENTATIONS:  USE    159␈↓

␈↓"β␈↓ α←␈↓␈↓ β?Design␈αB␈αis␈αhow␈αthe␈α␈↓	RELATIONS␈↓␈αare␈αdesigned.␈α In␈αthis␈αcase␈αthe␈α
traffic
␈↓ α←␈↓director␈α
can␈α
decide␈α
exactly␈α
which␈α∞target(s)␈α
the␈α
new␈α
site␈α
should␈α
be␈α∞added␈α
to,
␈↓ α←␈↓but␈α∞it␈α∞does␈α∞not␈α∞know␈α
how␈α∞to␈α∞add␈α∞it␈α∞there.␈α
 This␈α∞time␈α∞it␈α∞would␈α∞say,␈α
``Here's
␈↓ α←␈↓where you belong, ask about how to be added when you get there.''
␈↓"β␈↓ α←␈↓␈↓ β?Now␈α⊃consider␈α⊃the␈α⊃advantages␈α⊃and␈α⊃difficulties␈α⊃associated␈α⊃with␈α⊂each
␈↓ α←␈↓alternative.
␈↓"β␈↓ α←␈↓␈↓ β?Design␈α⊂A␈α∂requires␈α⊂that␈α∂the␈α⊂targets␈α⊂(or␈α∂the␈α⊂expert)␈α∂must␈α⊂be␈α⊂sure␈α∂to
␈↓ α←␈↓maintain␈α⊃the␈α⊃necessary␈α⊂constraints.␈α⊃ In␈α⊃terms␈α⊂of␈α⊃the␈α⊃current␈α⊃example,␈α⊂this
␈↓ α←␈↓would␈α∞mean␈α
either␈α∞including␈α∞in␈α
the␈α∞updating␈α∞function␈α
for␈α∞each␈α∞category␈α
of
␈↓ α←␈↓site␈α∂a␈α∂test␈α∂to␈α∞insure␈α∂mutual␈α∂exclusion␈α∂with␈α∞the␈α∂other␈α∂two␈α∂categories␈α∞(which
␈↓ α←␈↓would␈α∀be␈α∪slow␈α∀and␈α∀redundant)␈α∪or,␈α∀when␈α∀asking␈α∪the␈α∀expert␈α∀about␈α∪each
␈↓ α←␈↓category␈α
of␈α
site␈αindividually,␈α
relying␈α
on␈α
him␈αto␈α
maintain␈α
the␈α
requirement␈αof
␈↓ α←␈↓mutual␈α∞exclusion␈α∞(which␈α
would␈α∞be␈α∞slow,␈α
redundant,␈α∞and␈α∞less␈α∞reliable).␈α
 The
␈↓ α←␈↓traffic␈α⊂director␈α∂in␈α⊂design␈α∂B␈α⊂has␈α∂enough␈α⊂information␈α∂to␈α⊂present␈α⊂the␈α∂expert
␈↓ α←␈↓with␈α
a␈α
single␈α
coherent␈α
picture␈α
of␈α∞the␈α
choice␈α
to␈α
be␈α
made␈α
(e.g.,␈α
asking␈α∞him␈α
to
␈↓ α←␈↓choose␈α∩just␈α∪one␈α∩of␈α∪the␈α∩three␈α∪alternatives),␈α∩rather␈α∪than␈α∩requiring␈α∪him␈α∩to
␈↓ α←␈↓reconstruct it from a sequence of questions.
␈↓"β␈↓ α←␈↓␈↓ β?Design␈α∞C␈α
has␈α∞the␈α
disadvantage␈α∞that␈α
adding␈α∞a␈α
new␈α∞representation␈α
to
␈↓ α←␈↓the␈α∩system␈α∩would␈α∩be␈α∩rather␈α∩involved,␈α∩since␈α∩describing␈α∩its␈α∩traffic␈α⊃director
␈↓ α←␈↓would␈αrequire␈αkeeping␈αin␈αmind␈αthe␈αstructure␈αof␈αeach␈αtarget.␈α In␈αaddition,␈α
any
␈↓ α←␈↓changes␈αin␈α
the␈αstructure␈α
of␈αthe␈αtargets␈α
would␈αbe␈α
harder␈αto␈αaccommodate␈α
since
␈↓ α←␈↓knowledge␈α∂about␈α∂that␈α∂structure␈α∂might␈α∂be␈α∂widely␈α∂distributed␈α∂among␈α∂several
␈↓ α←␈↓traffic␈α
directors.␈α With␈α
alternative␈αB,␈α
all␈αthe␈α
necessary␈αchanges␈α
can␈α
be␈αmade
␈↓ α←␈↓by editing a single schema.
␈↓"β␈↓ α←␈↓␈↓ β?Design␈αB␈αhas␈αthe␈αadvantage␈αthat␈αall␈αof␈αthe␈αinformation␈αrelevant␈αto␈αa
␈↓ α←␈↓representation␈α↔is␈α↔associated␈α↔directly␈α↔with␈α↔it.␈α↔ This␈α↔offers␈α↔a␈α⊗convenient
␈↓ α←␈↓framework␈α
for␈αits␈α
initial␈α
acquisition␈αand␈α
can␈α
mean␈αmodifications␈α
are␈αeasier␈α
to
␈↓ α←␈↓make.␈α In␈αterms␈αof␈αthese␈αalternatives,␈α
it␈αcan␈αbe␈αseen␈αas␈αa␈αcompromise␈α
between
␈↓ α←␈↓having␈α∞the␈α∞information␈α∂associated␈α∞primarily␈α∞with␈α∂the␈α∞trigger␈α∞(i.e.,␈α∂stored␈α∞in
␈↓ α←␈↓the␈α∞traffic␈α∂director,␈α∞as␈α∂in␈α∞alternative␈α∂C)␈α∞and␈α∂having␈α∞it␈α∂associated␈α∞primarily
␈↓ α←␈↓with␈α∞the␈α∞targets␈α∞(i.e.,␈α∞stored␈α∞in␈α∞the␈α∞updating␈α∞functions,␈α∞as␈α∞in␈α∞alternative␈α
A).
␈↓ α←␈↓In␈αthe␈α
most␈αgeneral␈α
terms,␈αDesign␈αB␈α
succeeds␈αbecause␈α
␈↓↓it␈αkeeps␈αthe␈α
distribution
␈↓ α←␈↓↓of␈α∪information␈α∪about␈α∀data␈α∪structures␈α∪constrained␈α∀to␈α∪the␈α∪fewest␈α∀number␈α∪of
␈↓ α←␈↓↓locations␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈α∂that␈α∂this␈α∂updating␈α∂technique␈α∂is␈α∂applicable␈α∂to␈α∂a␈α∂wide␈α∂range␈α∞of
␈↓ α←␈↓data␈α⊃structures.␈α⊃ ␈↓	SITE-INFECT␈↓,␈α⊃for␈α⊃instance,␈α⊂is␈α⊃a␈α⊃table␈α⊃with␈α⊃a␈α⊃culture␈α⊂site
␈↓ α←␈↓labeling␈αeach␈α
row␈αand␈α
an␈αorganism␈α
identity␈αlabeling␈α
each␈αcolumn.␈α The␈α
entry
␈↓ α←␈↓in␈α∂that␈α∂row␈α∂and␈α∂column␈α∂is␈α∂the␈α∂CF␈α∂that␈α∂an␈α∂infection␈α∂at␈α∂site␈α⊂<rowname>␈α∂is
␈↓ α←␈↓caused␈α→by␈α→the␈α→organism␈α~<columnname>.␈α→ A␈α→newly␈α→acquired␈α~site␈α→will
␈↓ α←␈↓eventually␈α∞be␈α
sent␈α∞to␈α
␈↓	SITE-INFECT␈↓␈α∞as␈α∞part␈α
of␈α∞the␈α
response␈α∞to␈α∞the␈α
updating
␈↓ α←␈↓command:

␈↓"β␈↓ α←␈↓	␈↓ ∧?(ADDTO (AND* ALLSITES SITE-INFECT))

␈↓ α←␈↓In␈α∞this␈α∞case,␈α∞␈↓	SITE-INFECT␈↓␈α
sends␈α∞a␈α∞request␈α∞to␈α∞the␈α
schema␈α∞of␈α∞which␈α∞it␈α∞is␈α
an
␈↓ α←␈↓␈↓160    KNOWLEDGE ACQUISITION II␈↓ 
#6-9␈↓

␈↓"β␈↓ α←␈↓instance;␈αthis␈αschema␈αthen␈αinvokes␈αits␈αupdating␈αfunction,␈αwhich␈αresults␈αin␈αthe
␈↓ α←␈↓interaction␈α∞seen␈α∞earlier␈α∂in␈α∞the␈α∞trace␈α∞(``␈↓	What␈α∂are␈α∞the␈α∞likely␈α∂pathogens␈α∞to
␈↓ α←␈↓	be␈αfound␈αat␈αthe␈αsite:␈αURINE?␈↓'').␈α
 The␈αanswer␈αis␈αused␈αto␈αcreate␈αa␈α
new␈αrow
␈↓ α←␈↓in the ␈↓	SITE-INFECT␈↓ table.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∞caveat␈α∞mentioned␈α∞above␈α∞should␈α∞be␈α∞reemphasized.␈α∂ The␈α∞current
␈↓ α←␈↓design␈α∃scheme␈α∀takes␈α∃advantage␈α∀of␈α∃a␈α∀degree␈α∃of␈α∀modularity␈α∃in␈α∃the␈α∀data
␈↓ α←␈↓structures.␈α∞ It␈α∞is␈α∞applicable␈α∞only␈α
where␈α∞target␈α∞updating␈α∞is␈α∞not␈α∞dependent␈α
on
␈↓ α←␈↓extensive␈α∞information␈α∞from␈α∞the␈α∂trigger.␈α∞ That␈α∞is,␈α∞the␈α∞updating␈α∂functions␈α∞of
␈↓ α←␈↓each␈α∂target␈α∂in␈α⊂Fig.␈α∂6-7␈α∂must␈α⊂be␈α∂able␈α∂to␈α∂add␈α⊂new␈α∂elements␈α∂to␈α⊂their␈α∂targets
␈↓ α←␈↓without␈α
knowing␈α
which␈α
traffic␈α
director␈α
sent␈α
them␈α
the␈α
new␈α
element.␈α Since␈α
this
␈↓ α←␈↓modularity␈α∂is␈α⊂not␈α∂present␈α⊂in␈α∂all␈α∂data␈α⊂structure␈α∂designs,␈α⊂it␈α∂forms␈α⊂a␈α∂limiting
␈↓ α←␈↓factor in the approach.

␈↓"β␈↓ α←␈↓␈↓αNoting the new instance␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈αfinal␈αstep␈αin␈α``interpreting''␈αa␈αschema␈αis␈αto␈αadd␈αthe␈αnewly␈αcreated
␈↓ α←␈↓structure␈αto␈αthe␈αlist␈αof␈α␈↓	INSTANCES␈↓␈αof␈αthe␈αschema.␈α This␈αis␈αdone␈αprimarily␈αfor
␈↓ α←␈↓bookkeeping␈αpurposes,␈αbut␈αit␈αalso␈αhas␈αother␈αuseful␈αapplications␈αwhich␈αwill␈α
be
␈↓ α←␈↓demonstrated later.

␈↓"β␈↓ α←␈↓␈↓α6-9-2    Where to start in the network␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αdescription␈α
of␈αthe␈α
use␈αof␈αschemata␈α
to␈αguide␈α
acquisition␈αassumed
␈↓ α←␈↓that␈αthe␈αquestion␈αof␈α
where␈αto␈αstart␈αin␈α
the␈αschema␈αhierarchy␈αhad␈αalready␈α
been
␈↓ α←␈↓settled.␈α
 While␈α
the␈α∞mechanisms␈α
used␈α
to␈α
make␈α∞this␈α
decision␈α
are␈α∞not␈α
complex,
␈↓ α←␈↓they illustrate an interesting issue.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∩mechanism␈α∩provides␈α∩a␈α∩default␈α∩starting␈α∩place␈α∩for␈α∩the␈α∩case␈α∩in
␈↓ α←␈↓which␈α
the␈α
user␈α
indicates,␈α
outside␈α
of␈α
the␈α
context␈α
of␈α
any␈α
consultation,␈α
that␈α
he
␈↓ α←␈↓wants␈αto␈α
teach␈αthe␈α
system␈αabout␈α
some␈αnew␈α
instance.␈α (While␈α
we␈αhave␈αseen␈α
the
␈↓ α←␈↓acquisition␈αof␈αa␈α
new␈αvalue␈αillustrated␈α
in␈αthe␈αcontext␈α
of␈αrule␈αacquisition,␈α
it␈αis
␈↓ α←␈↓also␈α∃possible␈α∀to␈α∃acquire␈α∀new␈α∃instances␈α∃of␈α∀any␈α∃data␈α∀type␈α∃as␈α∃a␈α∀separate
␈↓ α←␈↓operation.)␈α
Since␈α
there␈αis␈α
no␈α
context␈α
to␈αrely␈α
on,␈α
the␈α
default␈αis␈α
to␈α
start␈α
at␈αthe
␈↓ α←␈↓root␈α
of␈α
the␈α
schema␈α
network␈α
and␈α
ask␈α
the␈α
expert␈α
to␈α
choose␈α
the␈α
path␈α∞at␈α
every
␈↓ α←␈↓branch␈α∞point.␈α∞ This␈α∞presents␈α∞a␈α∞reasonable␈α∞dialog␈α∞since␈α∞it␈α∞requests␈α∂from␈α∞the
␈↓ α←␈↓expert␈α∞a␈α∂progressively␈α∞more␈α∂detailed␈α∞specification␈α∂of␈α∞the␈α∂concept␈α∞he␈α∂has␈α∞in
␈↓ α←␈↓mind.␈α Each␈αindividual␈αinquiry␈αwill␈αappear␈αsensible␈αsince,␈αwithout␈α
contextual
␈↓ α←␈↓information,␈αthere␈αis␈αno␈αway␈αthe␈αsystem␈αcould␈αhave␈αdeduced␈αthe␈αanswer.␈α (In
␈↓ α←␈↓the␈α∞excerpt␈α∞below,␈α∞only␈α∞the␈α∞sequence␈α∞of␈α∞questions␈α∞is␈α∞shown;␈α∞everything␈α
else
␈↓ α←␈↓has been edited out.)

␈↓"β␈↓ α←␈↓	++** ␈↓α?␈↓	
␈↓"β␈↓ α←␈↓	Commands are
␈↓"β␈↓ α←␈↓	        NR - enter a new rule
␈↓"β␈↓ α←␈↓	        ER - edit an existing rule
␈↓"β␈↓ α←␈↓	        DR - delete rule
␈↓"β␈↓ α←␈↓	        NP - enter a new primitive (attribute, value etc.)
␈↓"β␈↓ α←␈↓	++** ␈↓αNP␈↓	

␈↓"β␈↓ α←␈↓	Which of the following best describes the new primitive?
␈↓"β␈↓ α←␈↓␈↓6-9␈↓ ¬XKNOWLEDGE ABOUT REPRESENTATIONS:  USE    161␈↓

␈↓"β␈↓ α←␈↓	[Choose the last if no other is appropriate]
␈↓"β␈↓ α←␈↓	     1 - an attribute, or
␈↓"β␈↓ α←␈↓	     2 - a value of an attribute, or
␈↓"β␈↓ α←␈↓	     3 - None of the above
␈↓"β␈↓ α←␈↓	Choose one
␈↓"β␈↓ α←␈↓	++**␈↓α 1␈↓	

␈↓ α←␈↓↓␈↓ βW{At␈αthis␈α
point,␈αacquisition␈α
of␈αthe␈αnew␈α
item␈αwould␈α
begin;␈αit␈α
is␈αomitted
␈↓ α←␈↓↓␈↓ βWhere.}

␈↓"β␈↓ α←␈↓	Which of the following best describes the new attribute?
␈↓"β␈↓ α←␈↓	[Choose the last if no other is appropriate]
␈↓"β␈↓ α←␈↓	     1 - an attribute of a patient
␈↓"β␈↓ α←␈↓	     2 - an attribute of a infection
␈↓"β␈↓ α←␈↓	     3 - an attribute of a culture
␈↓"β␈↓ α←␈↓	     4 - an attribute of a organism
␈↓"β␈↓ α←␈↓	     5 - None of the above
␈↓"β␈↓ α←␈↓	Choose one
␈↓"β␈↓ α←␈↓	++**␈↓α3␈↓	

␈↓ α←␈↓↓␈↓ βW{Here␈α∂we␈α∂would␈α∂see␈α∂additional␈α∂acquisition␈α∂of␈α∂information␈α∂about␈α∞the
␈↓ α←␈↓↓␈↓ βWitem; again, omitted.}

␈↓"β␈↓ α←␈↓	Which of the following best describes the new attribute?
␈↓"β␈↓ α←␈↓	[Choose the last if no other is appropriate]
␈↓"β␈↓ α←␈↓	     1 - a single-valued attribute, or
␈↓"β␈↓ α←␈↓	     2 - a multi-valued attribute, or
␈↓"β␈↓ α←␈↓	     3 - an attribute whose value is "true" or "false", or
␈↓"β␈↓ α←␈↓	     4 - None of the above
␈↓"β␈↓ α←␈↓	Choose one
␈↓"β␈↓ α←␈↓	++**␈↓α 3␈↓	

␈↓"β␈↓ α←␈↓␈↓ β?When␈αa␈α
new␈αconcept␈α
is␈αmentioned␈α
during␈αa␈α
rule␈αacquisition,␈α
however,
␈↓ α←␈↓there␈αis␈αan␈αextensive␈αamount␈αof␈αcontext␈αavailable.␈α The␈αsame␈αsort␈αof␈αdefault
␈↓ α←␈↓approach␈α⊂would␈α⊂look␈α⊂``dumb''␈α⊂in␈α∂this␈α⊂case,␈α⊂since␈α⊂there␈α⊂are␈α⊂numerous␈α∂clues
␈↓ α←␈↓indicating␈α∞which␈α∞kind␈α∞of␈α
data␈α∞type␈α∞is␈α∞being␈α
mentioned.␈α∞ In␈α∞the␈α∞example␈α
in
␈↓ α←␈↓Section␈α∞6-7-1,␈α∞for␈α∞instance,␈α
it␈α∞was␈α∞not␈α∞difficult␈α
to␈α∞discover␈α∞that␈α∞the␈α
concept
␈↓ α←␈↓was␈α
a␈α
new␈α
identity␈α
of␈α
an␈α
organism.␈α
 As␈α
was␈α
indicated,␈α
this␈α
is␈αaccomplished␈α
by
␈↓ α←␈↓some␈αsimple␈αpattern␈αmatching.␈α
 Each␈αschema␈αin␈αthe␈α
network␈αhas␈αone␈αor␈α
more
␈↓ α←␈↓patterns associated with it.  For example, the pattern

␈↓"β␈↓ α←␈↓	␈↓ ∧Othe <attribute> of <object> is --

␈↓ α←␈↓is␈αassociated␈αwith␈αthe␈α␈↓	VALUE-SCHEMA␈↓.␈α Each␈αof␈αthese␈αis␈αtested␈αagainst␈αthe␈α
line
␈↓ α←␈↓of␈αtext␈αthat␈αprompted␈αacquisition␈αof␈αthe␈αnew␈αitem,␈αand␈αthe␈αoutcome␈αsupplies
␈↓ α←␈↓a␈α⊃starting␈α⊃place␈α⊃in␈α⊂the␈α⊃network.␈α⊃ (If␈α⊃all␈α⊂matches␈α⊃fail,␈α⊃the␈α⊃system␈α⊃starts␈α⊂as
␈↓ α←␈↓before with the root of the network.)
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
patterns␈α
thus␈α
make␈α
it␈α
possible␈α
to␈α
use␈α
contextual␈αinformation␈α
from
␈↓ α←␈↓the␈α∞rule␈α∞acquisition␈α
dialog␈α∞in␈α∞order␈α
to␈α∞select␈α∞a␈α
starting␈α∞place␈α∞in␈α∞the␈α
schema
␈↓"β␈↓ α←␈↓␈↓162    KNOWLEDGE ACQUISITION II␈↓ 
#6-9␈↓

␈↓"β␈↓ α←␈↓network.␈α∂ Note␈α∞that␈α∂this␈α∂link␈α∞between␈α∂the␈α∂natural␈α∞language␈α∂dialog␈α∂and␈α∞the
␈↓ α←␈↓data␈αtype␈αhierarchy␈αrepresents␈αpart␈αof␈αthe␈αsemantics␈αof␈αeach␈αdata␈αtype.␈α
 Since
␈↓ α←␈↓the␈αschemata␈αwere␈αdesigned␈αinitially␈αto␈αrepresent␈αonly␈αthe␈αsyntax␈αof␈αthe␈αdata
␈↓ α←␈↓types,␈α∂at␈α∂present␈α⊂they␈α∂contain␈α∂only␈α∂the␈α⊂very␈α∂limited␈α∂and␈α∂somewhat␈α⊂ad␈α∂hoc
␈↓ α←␈↓semantic␈α⊂information␈α⊂in␈α⊂the␈α⊂patterns.␈α⊂ Such␈α⊂information␈α⊂is␈α⊃clearly␈α⊂needed,
␈↓ α←␈↓however,␈α
and␈αwould␈α
represent␈αa␈α
useful␈αand␈α
natural␈αextension␈α
to␈α
the␈αcurrent
␈↓ α←␈↓implementation.␈α
 It␈α
would␈α
mean␈α
that,␈α
along␈α
with␈α
the␈α
syntax␈α
of␈α
each␈α
data␈α
type,
␈↓ α←␈↓some␈α∂of␈α∂its␈α∂semantics␈α⊂would␈α∂be␈α∂described,␈α∂perhaps␈α⊂in␈α∂the␈α∂form␈α∂of␈α⊂a␈α∂more
␈↓ α←␈↓systematic␈α↔set␈α⊗of␈α↔patterns␈α↔than␈α⊗those␈α↔currently␈α↔in␈α⊗use,␈α↔or␈α↔other␈α⊗more
␈↓ α←␈↓sophisticated␈αdevices.␈α The␈αsystem␈αwould␈αthen␈αalways␈αstart␈αat␈αthe␈αroot␈αof␈αthe
␈↓ α←␈↓network␈αand␈αcould␈αuse␈αthe␈α
semantic␈αinformation␈αstored␈αwith␈αeach␈α
schema␈αto
␈↓ α←␈↓take␈α
advantage␈α
of␈αcontext␈α
from␈α
the␈α
dialog,␈αguiding␈α
its␈α
own␈α
descent␈αthrough
␈↓ α←␈↓the network.

␈↓"β␈↓ α←␈↓␈↓α6-9-3    Schema function:  Access and storage␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
schema␈αconcept␈α
was␈αintroduced␈α
by␈αdescribing␈α
it␈αas␈α
an␈αextension
␈↓ α←␈↓of␈α⊂the␈α⊂notion␈α⊂of␈α⊂a␈α⊂record␈α⊃structure,␈α⊂and␈α⊂we␈α⊂have␈α⊂seen␈α⊂how␈α⊂it␈α⊃guides␈α⊂the
␈↓ α←␈↓acquisition␈αof␈αa␈αnew␈αinstance.␈α A␈α
second␈αimportant␈αuse␈αof␈αrecord␈αstructures␈α
is
␈↓ α←␈↓for␈α⊃access␈α⊂and␈α⊃storage,␈α⊃and␈α⊂the␈α⊃schemata␈α⊃have␈α⊂the␈α⊃parallel␈α⊃capability.␈α⊂ If
␈↓ α←␈↓slotnames␈α∀are␈α∀viewed␈α∃as␈α∀analogous␈α∀to␈α∀the␈α∃fields␈α∀of␈α∀a␈α∀record,␈α∃then␈α∀the
␈↓ α←␈↓mechanism␈αused␈α
in␈α␈↓¬TEIRESIAS␈↓␈αlooks␈α
quite␈αsimilar␈αto␈α
the␈αstandard␈α
record␈α␈↓↓fetch␈↓
␈↓ α←␈↓and ␈↓↓store␈↓ operations.
␈↓"β␈↓ α←␈↓␈↓ β?Our␈αapproach␈α
is␈αbased␈αon␈α
generalizing␈αthe␈αuse␈α
of␈αthe␈α␈↓↓advice␈↓␈α
construct
␈↓ α←␈↓by␈α⊃using␈α⊃four␈α⊂additional␈α⊃types␈α⊃of␈α⊂advice␈α⊃(shown␈α⊃below).␈α⊂ To␈α⊃carry␈α⊃out␈α⊂a
␈↓ α←␈↓storage␈α
or␈αaccess␈α
operation,␈αthe␈α
relevant␈αslotexpert␈α
is␈αsent␈α
the␈αthe␈α
name␈αof␈α
the
␈↓ α←␈↓data structure and one of these pieces of advice.

␈↓"β␈↓ α←␈↓	  GETONE  ␈↓retrieve one instance of whatever fills this slot in the␈↓	
␈↓"β␈↓ α←␈↓	          ␈↓indicated data structure␈↓	
␈↓"β␈↓ α←␈↓	  GETALL  ␈↓retrieve all instances of whatever fills this slot in the␈↓	
␈↓"β␈↓ α←␈↓	          ␈↓indicated data structure␈↓	
␈↓"β␈↓ α←␈↓	  GETNEXT ␈↓retrieve one new instance at each request␈↓	
␈↓"β␈↓ α←␈↓	  STOREIT ␈↓store an item in the slot of the indicated data structure

␈↓"β␈↓ α←␈↓α␈↓ β-Fig. 6-8.    Four pieces of advice used for access and storage.    


␈↓"β␈↓ α←␈↓␈↓ β?For instance

␈↓"β␈↓ α←␈↓	␈↓ ∧↔(APPLY* (GETEXPERT AIR) 'E.COLI 'GETONE)

␈↓ α←␈↓will retrieve one of the aerobicity values of ␈↓	E.COLI␈↓, while

␈↓"β␈↓ α←␈↓	␈↓ ∧∂(APPLY* (GETEXPERT AIR) 'E.COLI 'GETNEXT)

␈↓ α←␈↓functions␈α∂as␈α∂a␈α∂generator␈α∂and␈α∂will␈α∂retrieve␈α∂them␈α∂all␈α∂one␈α∂by␈α∂one.␈α∂ Since␈α∂the
␈↓ α←␈↓␈↓6-9␈↓ ¬XKNOWLEDGE ABOUT REPRESENTATIONS:  USE    163␈↓

␈↓"β␈↓ α←␈↓slotexperts␈α
are␈α
organized␈α
around␈α
the␈α∞pieces␈α
of␈α
advice,␈α
the␈α
relevant␈α∞code␈α
for
␈↓ α←␈↓storage or retrieval is similarly organized in each slotexpert.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α⊃feature,␈α⊂too,␈α⊃has␈α⊂been␈α⊃influenced␈α⊂by␈α⊃the␈α⊂work␈α⊃on␈α⊃actors␈α⊂and
␈↓ α←␈↓␈↓¬SMALLTALK␈↓␈α
noted␈α
earlier.␈α
 However,␈α∞where␈α
that␈α
work␈α
concentrates␈α
on␈α∞issues␈α
of
␈↓ α←␈↓programming␈α∪and␈α∪program␈α∩correctness,␈α∪we␈α∪intend␈α∩here␈α∪nothing␈α∪quite␈α∩so
␈↓ α←␈↓formidable.␈α
 We␈α∞use␈α
it␈α∞because␈α
it␈α
was␈α∞a␈α
natural␈α∞extension␈α
to␈α∞our␈α
approach,
␈↓ α←␈↓with much the same perspective on organization of knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?Our␈α∀implementation␈α∀offers,␈α∃for␈α∀instance,␈α∀the␈α∃well-known␈α∀benefits
␈↓ α←␈↓supplied␈α⊂by␈α⊂any␈α⊂record-like␈α⊂structure␈α∂that␈α⊂provides␈α⊂a␈α⊂level␈α⊂of␈α∂``insulation''
␈↓ α←␈↓between␈α∪representation␈α∪and␈α∀implementation.␈α∪ The␈α∪use␈α∪of␈α∀slotexperts␈α∪and
␈↓ α←␈↓advice␈α
turns␈α
accessing␈αa␈α
structure␈α
into␈αa␈α
process␈α
of␈αsending␈α
a␈α
request␈α
to␈αthe
␈↓ α←␈↓data␈α∩structure␈α∪itself,␈α∩which␈α∪then␈α∩``answers''␈α∩by␈α∪providing␈α∩(or␈α∪storing)␈α∩the
␈↓ α←␈↓desired␈αitem.␈α All␈αaccess␈αand␈αstorage␈αis␈αthus␈αfunneled␈αthrough␈αthe␈αindividual
␈↓ α←␈↓structures␈α↔(via␈α↔the␈α_slotexperts),␈α↔and␈α↔explicit␈α_reference␈α↔is␈α↔made␈α_to␈α↔the
␈↓ α←␈↓configuration␈α∩of␈α⊃the␈α∩structure␈α⊃in␈α∩only␈α∩one␈α⊃place␈α∩in␈α⊃the␈α∩system.␈α∩ As␈α⊃with
␈↓ α←␈↓standard␈α
record␈α
structures,␈α∞this␈α
technique␈α
makes␈α
it␈α∞possible␈α
to␈α
access␈α∞a␈α
data
␈↓ α←␈↓structure␈α⊂without␈α⊂reference␈α⊂to␈α⊂the␈α⊃details␈α⊂of␈α⊂how␈α⊂it␈α⊂is␈α⊂actually␈α⊃stored␈α⊂and
␈↓ α←␈↓without␈α∃the␈α∃need␈α∃to␈α∃change␈α∃the␈α∃code␈α∃if␈α∃the␈α∃storage␈α∃implementation␈α∀is
␈↓ α←␈↓modified.␈α∪ In␈α∪addition,␈α∪the␈α∪slotexperts␈α∪make␈α∪it␈α∪easy␈α∪to␈α∪use␈α∪an␈α∪arbitrary
␈↓ α←␈↓function␈α⊃for␈α⊃storage␈α⊃and␈α⊃retrieval.␈α∩ Dates,␈α⊃for␈α⊃example,␈α⊃are␈α⊃stored␈α∩in␈α⊃the
␈↓ α←␈↓system␈α∃as␈α∃integers␈α∃(for␈α∃efficiency),␈α∃and␈α∃the␈α∃␈↓	DATE-EXPERT␈↓␈α∃takes␈α⊗care␈α∃of
␈↓ α←␈↓decoding and encoding them on access and storage.
␈↓ α←␈↓␈↓164    KNOWLEDGE ACQUISITION II␈↓ 
#6-9␈↓

␈↓"β␈↓ α←␈↓␈↓α6-10    TRACE OF SYSTEM PERFORMANCE:  ACQUIRING A NEW
␈↓ α←␈↓α␈↓ β3ATTRIBUTE␈↓
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∞more␈α∞sophisticated␈α∂example--the␈α∞acquisition␈α∞of␈α∞a␈α∂new␈α∞attribute--
␈↓ α←␈↓will␈α⊗illustrate␈α⊗several␈α⊗other␈α⊗aspects␈α⊗of␈α⊗our␈α⊗approach␈α⊗to␈α⊗handling␈α⊗data
␈↓ α←␈↓structures.␈α It␈αwill␈α
demonstrate,␈αfor␈αinstance,␈αthe␈α
utility␈αof␈αthe␈αschema␈α
network
␈↓ α←␈↓as␈α∞a␈α
device␈α∞for␈α
structuring␈α∞the␈α
acquisition␈α∞process.␈α
 The␈α∞network␈α
is␈α∞used␈α
to
␈↓ α←␈↓organize␈α∀the␈α∀dialog␈α∀and␈α∀to␈α∃insure␈α∀that␈α∀the␈α∀expert␈α∀is␈α∀presented␈α∃with␈α∀a
␈↓ α←␈↓comprehensible␈αsequence␈αof␈αquestions.␈α It␈α
also␈αoffers␈αa␈αfoundation␈αfor␈α
adding
␈↓ α←␈↓new␈α∂data␈α∞structures␈α∂to␈α∂the␈α∞system,␈α∂making␈α∂the␈α∞task␈α∂reasonably␈α∂simple␈α∞both
␈↓ α←␈↓conceptually and computationally.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈αan␈α
attribute␈αis␈α
a␈αmore␈α
complex␈αstructure␈α
than␈αthose␈α
encountered
␈↓ α←␈↓in␈α_previous␈α↔examples,␈α_the␈α↔acquisition␈α_process␈α↔is␈α_correspondingly␈α↔more
␈↓ α←␈↓complex.␈α∂ In␈α∂particular,␈α∂the␈α∂presence␈α∂of␈α∂several␈α∂different␈α∂data␈α∂types␈α∂in␈α∂the
␈↓ α←␈↓substructure␈αof␈αthe␈αattribute␈αguides␈αthe␈αdialog␈αthrough␈αa␈αparallel␈αsequence␈α
of
␈↓ α←␈↓several different topics.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃example␈α⊃presents␈α⊃a␈α∩borderline␈α⊃case␈α⊃for␈α⊃the␈α∩distinction␈α⊃drawn
␈↓ α←␈↓earlier␈α∀between␈α∀formalism␈α∀and␈α∀degree␈α∪of␈α∀expertise.␈α∀ To␈α∀describe␈α∀a␈α∪new
␈↓ α←␈↓attribute␈α⊃it␈α⊃is␈α⊃necessary␈α⊃to␈α⊃describe␈α⊂the␈α⊃values␈α⊃associated␈α⊃with␈α⊃it.␈α⊃ It␈α⊂may
␈↓ α←␈↓happen␈α
that␈αthese␈α
values␈α
require␈αa␈α
data␈αtype␈α
as␈α
yet␈αunknown␈α
to␈α
the␈αsystem,
␈↓ α←␈↓and␈α∀creating␈α∀that␈α∀data␈α∀type␈α∪becomes,␈α∀in␈α∀turn,␈α∀a␈α∀new␈α∀subproblem.␈α∪ But
␈↓ α←␈↓describing␈αa␈αnew␈αkind␈αof␈αdata␈αtype␈αis␈αa␈αcomplex␈αoperation␈αthat␈αtypically␈αhas
␈↓ α←␈↓an␈α∞impact␈α∞on␈α∂the␈α∞underlying␈α∞formalism␈α∂of␈α∞the␈α∞performance␈α∂program.␈α∞ The
␈↓ α←␈↓current␈αexample␈αis␈αa␈α
borderline␈αcase␈αbecause␈αthat␈α
impact␈αcan␈αbe␈αavoided,␈α
but
␈↓ α←␈↓it will demonstrate how the problem arises.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αdialog␈αbelow␈αis␈αpresented␈αin␈αtwo␈αparts.␈α After␈αthe␈αfirst␈αhalf␈αthere
␈↓ α←␈↓is␈αa␈αbrief␈αreview␈α
of␈αwhat␈αhas␈αhappened␈αso␈α
far␈αand␈αan␈αindication␈αof␈α
what␈αis
␈↓ α←␈↓yet␈αto␈αoccur.␈α Recall␈αthat␈αthe␈αknowledge␈αbase␈αappears␈αas␈αit␈αwould␈αduring␈αthe
␈↓ α←␈↓early stages of construction.

␈↓ α←␈↓	***************************************************************

␈↓ α←␈↓	The new rule will be called RULE383
␈↓ α←␈↓	 If:    1 -␈↓α  THE ORGANISM IS A GRAM NEGATIVE ROD␈↓	
␈↓ α←␈↓	   and  2 -␈↓α  PATIENT IS A COMPROMISED HOST␈↓	
␈↓ α←␈↓	   and  3 -␈↓α  THE NUTRIENT OF THE CULTURE MEDIUM IS␈↓	
␈↓ α←␈↓	           ␈↓α  BLOOD-AGAR␈↓	
␈↓ α←␈↓	   and  4 -
␈↓ α←␈↓	 Then:  1 -␈↓α  THERE IS EVIDENCE (.8) THAT THE CATEGORY IS␈↓	
␈↓ α←␈↓	           ␈↓α  ENTEROBACTERIACEAE␈↓	
␈↓ α←␈↓	   and  2 -

␈↓ α←␈↓	This may take a few moments...

␈↓ α←␈↓	I don't understand
␈↓ α←␈↓	          THE NUTRIENT OF THE CULTURE MEDIUM IS BLOOD-AGAR
␈↓ α←␈↓	because these words are unknown: NUTRIENT MEDIUM BLOOD-AGAR
␈↓ α←␈↓␈↓6-10␈↓ π∂ACQUIRING A NEW ATTRIBUTE    165␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βW{The␈αexpert␈αis␈αtyping␈αin␈αa␈αnew␈αrule␈αand␈αuses␈αseveral␈αwords␈α␈↓¬TEIRESIAS␈↓↓
␈↓ α←␈↓↓␈↓ βWdoesn't␈αrecognize.␈α As␈α
before,␈αthe␈αsystem␈α
indicates␈αits␈αinability␈αto␈α
parse
␈↓ α←␈↓↓␈↓ βWone of his lines and displays the unfamiliar words.}

␈↓ α←␈↓	Would you care to try rephrasing that?
␈↓ α←␈↓	++** ␈↓αY␈↓	
␈↓ α←␈↓	  (Please use *'s to mark what you think might be a new
␈↓ α←␈↓	   attribute or a new value of an attribute)
␈↓ α←␈↓	        3 -  ␈↓α* THE NUTRIENT * OF THE CULTURE MEDIUM IS␈↓	
␈↓ α←␈↓	             ␈↓α  BLOOD-AGAR␈↓	
␈↓ α←␈↓	It looks as though
␈↓ α←␈↓	     THE NUTRIENT
␈↓ α←␈↓	refers to a new attribute, correct?
␈↓ α←␈↓	++**␈↓α Y␈↓	
␈↓ α←␈↓	 ...is it a new attribute of the culture?
␈↓ α←␈↓	++**␈↓α Y␈↓	

␈↓ α←␈↓↓␈↓ βW{Once␈α∃again,␈α⊗the␈α∃language␈α⊗is␈α∃sufficiently␈α⊗simple␈α∃that␈α⊗with␈α∃the
␈↓ α←␈↓↓␈↓ βWrelevant␈α~phrase␈α≠marked␈α~off,␈α~straightforward␈α≠pattern␈α~matching
␈↓ α←␈↓↓␈↓ βWsucceeds␈α∪in␈α∀determining␈α∪what␈α∀the␈α∪new␈α∪item␈α∀is.␈α∪The␈α∀system␈α∪then
␈↓ α←␈↓↓␈↓ βWretrieves the schema for attributes and begins to fill it out.}

␈↓ α←␈↓	  Now tell me a few things about this new attribute ...

␈↓ α←␈↓	  If there is one word which neatly sums up "THE NUTRIENT",
␈↓ α←␈↓	  please give
␈↓ α←␈↓	  it (else press RETURN)
␈↓ α←␈↓	  ++**␈↓α NUTRIENT␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈αnew␈αdata␈αstructure␈αneeds␈αa␈αname,␈αso␈αthe␈αsystem␈αasks␈αthe␈αexpert
␈↓ α←␈↓↓␈↓ βWto␈α⊂supply␈α⊃one.␈α⊂If␈α⊃he␈α⊂declines,␈α⊂it␈α⊃will␈α⊂create␈α⊃one,␈α⊂using␈α⊃the␈α⊂phrase
␈↓ α←␈↓↓␈↓ βWdelimited␈α∂by␈α∂the␈α⊂asterisks␈α∂and␈α∂a␈α⊂few␈α∂simple␈α∂heuristics␈α⊂for␈α∂creating
␈↓ α←␈↓↓␈↓ βWmnemonic variable names.}

␈↓ α←␈↓	  If someone said WHAT DO YOU MEAN BY NUTRIENT, what would you
␈↓ α←␈↓	  say?
␈↓ α←␈↓	  [type an empty line when done]
␈↓ α←␈↓	  ++**␈↓α CULTURE MEDIA COME WITH A WHOLE RANGE OF␈↓	
␈↓ α←␈↓	  ++**␈↓α NUTRIENTS, AND SOME BACTERIA ARE VERY PARTIAL␈↓	
␈↓ α←␈↓	  ++**␈↓α TO ONE NUTRIENT OR ANOTHER␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{As␈α
before,␈α∞the␈α
author␈α
of␈α∞a␈α
new␈α
concept␈α∞is␈α
responsible␈α∞for␈α
describing
␈↓ α←␈↓↓␈↓ βWit.}

␈↓ α←␈↓	  Please give a phrase with which I can talk about the value
␈↓ α←␈↓	  of NUTRIENT
␈↓ α←␈↓	  ++**␈↓α THE NUTRIENT OF THE MEDIUM OF *␈↓	
␈↓"β␈↓ α←␈↓␈↓166    KNOWLEDGE ACQUISITION II␈↓ 
∪6-10␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βW{All␈αattributes␈α
have␈αa␈α
phrase␈αassociated␈α
with␈αthem␈α
that␈αcan␈α
be␈αused␈α
to
␈↓ α←␈↓↓␈↓ βWtalk␈α⊃about␈α⊃their␈α⊂value.␈α⊃ It␈α⊃is␈α⊃used␈α⊂in␈α⊃many␈α⊃places␈α⊃throughout␈α⊂the
␈↓ α←␈↓↓␈↓ βWsystem␈α∂(e.g.,␈α∂to␈α∂produce␈α∂the␈α∂natural␈α∂language␈α∂version␈α∂of␈α⊂the␈α∂rules).
␈↓ α←␈↓↓␈↓ βWThe␈α∩asterisk␈α∩is␈α⊃filled␈α∩in␈α∩with␈α⊃the␈α∩name␈α∩of␈α⊃the␈α∩relevant␈α∩object,␈α⊃to
␈↓ α←␈↓↓␈↓ βWproduce, for instance, ``the nutrient of the medium of CULTURE-1.''}

␈↓ α←␈↓	      Now we want to select some keywords which might be good
␈↓ α←␈↓	      indicators that a speaker was referring to this new
␈↓ α←␈↓	      attribute. There are two classes of keywords (strong and
␈↓ α←␈↓	      weak) that are used to classify each of the non-trivial
␈↓ α←␈↓	      words in the phrase you just typed.
␈↓ α←␈↓	      Should NUTRIENT be a keyword?
␈↓ α←␈↓	      ++**␈↓α Y␈↓	
␈↓ α←␈↓	      Strong?
␈↓ α←␈↓	      ++**␈↓α Y␈↓	
␈↓ α←␈↓	      Should MEDIUM be a keyword?
␈↓ α←␈↓	      ++**␈↓α Y␈↓	
␈↓ α←␈↓	      Strong?
␈↓ α←␈↓	      ++**␈↓α Y␈↓	
␈↓ α←␈↓	      If you want to add any additional keywords,type them now.
␈↓ α←␈↓	      [type an empty line when done]
␈↓ α←␈↓	      ++**␈↓α ENERGY-SOURCE␈↓	
␈↓ α←␈↓	      Strong?
␈↓ α←␈↓	      ++**␈↓α N␈↓	
␈↓ α←␈↓	      ++**

␈↓ α←␈↓↓␈↓ βW{As␈α
described␈α
in␈α
chapter␈α
5,␈α
natural␈α
language␈α
capabilities␈α
are␈αbased␈α
on
␈↓ α←␈↓↓␈↓ βWkeywords␈α∞selected␈α∞from␈α∞the␈α∞English␈α∞phrases␈α∞associated␈α∞with␈α∂some␈α∞of
␈↓ α←␈↓↓␈↓ βWthe␈α∃data␈α∃types.␈α∃ Here␈α⊗the␈α∃dictionary␈α∃is␈α∃updated␈α∃to␈α⊗include␈α∃the
␈↓ α←␈↓↓␈↓ βWkeywords from the new attribute.}

␈↓ α←␈↓	  Is this attribute the result of a quantitative laboratory
␈↓ α←␈↓	  procedure?
␈↓ α←␈↓	  ++**␈↓α Y␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈αanswer␈αwill␈αbe␈α
important␈αduring␈αconsultations: ␈αRecall␈α
that␈αfor
␈↓ α←␈↓↓␈↓ βWthis␈α
type␈α
of␈αattribute,␈α
the␈α
consultation␈α
system␈αasks␈α
for␈α
the␈αvalue␈α
before
␈↓ α←␈↓↓␈↓ βWattempting to deduce it.}

␈↓ α←␈↓	Which of the following best describes concepts like NUTRIENT
␈↓ α←␈↓	[Choose the last if no other is appropriate]
␈↓ α←␈↓	     1 - a single-valued attribute, or
␈↓ α←␈↓	     2 - a multi-valued attribute, or
␈↓ α←␈↓	     3 - an attribute whose valued is "true" or "false", or
␈↓ α←␈↓	     4 - None of the above
␈↓ α←␈↓	Choose one
␈↓ α←␈↓	++**␈↓α 1␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈α↔expert␈α_has␈α↔supplied␈α↔all␈α_the␈α↔information␈α↔required␈α_by␈α↔the
␈↓"β␈↓ α←␈↓␈↓6-10␈↓ π∂ACQUIRING A NEW ATTRIBUTE    167␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βW␈↓	CULATTRIB-SCHEMA␈↓↓,␈α∂and␈α∂␈↓¬TEIRESIAS␈↓↓␈α∞now␈α∂attempts␈α∂to␈α∂descend␈α∞further
␈↓ α←␈↓↓␈↓ βWdown␈αthe␈α
schema␈αnetwork.␈α
 As␈αFig.␈α6-3␈α
illustrated,␈αhowever,␈α
there␈αis
␈↓ α←␈↓↓␈↓ βWa␈α∂three-way␈α∞branch␈α∂here.␈α∂ Since␈α∞the␈α∂system␈α∂has␈α∞no␈α∂way␈α∂of␈α∞knowing
␈↓ α←␈↓↓␈↓ βWwhich␈α⊗way␈α⊗to␈α⊗go,␈α⊗it␈α∃asks␈α⊗the␈α⊗expert,␈α⊗who␈α⊗responds␈α⊗by␈α∃further
␈↓ α←␈↓↓␈↓ βWspecifying the concept he has in mind.}

␈↓ α←␈↓	  Please give all the legal values for NUTRIENT
␈↓ α←␈↓	  The answer should be in the form of 1 or more of the
␈↓ α←␈↓	  following-   <a value of an attribute>
␈↓ α←␈↓	  ++**␈↓α BLOOD-AGAR␈↓	
␈↓ α←␈↓	  ++**␈↓α THAYER-MARTIN␈↓	
␈↓ α←␈↓	  ++**

␈↓ α←␈↓↓␈↓ βW{One␈α∀of␈α∀the␈α∀components␈α∀of␈α∃an␈α∀attribute␈α∀is␈α∀a␈α∀list␈α∀of␈α∃the␈α∀values
␈↓ α←␈↓↓␈↓ βWassociated␈αwith␈αit.␈α This␈αis␈α
indicated␈αin␈αthe␈αrelevant␈αschema␈α
with␈αthe
␈↓ α←␈↓↓␈↓ βWslotname-blank-advice triple.
␈↓ α←␈↓	␈↓ βWLEGALVALUES (KLEENE (1) < VALUE-INST >) ASKIT
␈↓ α←␈↓	␈↓ βWwhich␈α⊗says␈α⊗that␈α⊗an␈α↔attribute␈α⊗has␈α⊗one␈α⊗or␈α↔more␈α⊗objects
␈↓ α←␈↓	␈↓ βWassociated␈α⊃with␈α⊃it␈α⊃that␈α⊃are␈α⊂of␈α⊃type␈α⊃VALUE␈α⊃and␈α⊃that␈α⊂they
␈↓ α←␈↓	␈↓ βWshould be obtained from the expert.}

␈↓ α←␈↓	  Now I need to know a few things about each of these values...
␈↓ α←␈↓	  First BLOOD-AGAR

␈↓ α←␈↓↓␈↓ βW{Since␈αthey␈αare␈αnew␈αto␈αthe␈αsystem,␈αthe␈αexpert␈αis␈αasked␈αto␈αdescribe␈α
each
␈↓ α←␈↓↓␈↓ βWof␈α∩them.␈α∪ Since␈α∩the␈α∩triple␈α∪shown␈α∩above␈α∩indicates␈α∪that␈α∩each␈α∪is␈α∩an
␈↓ α←␈↓↓␈↓ βWinstance␈α⊂of␈α⊂the␈α⊂␈↓	VALUE␈↓↓␈α⊂schema,␈α⊂the␈α⊂system␈α⊂starts␈α⊂by␈α⊂retrieving␈α⊂that
␈↓ α←␈↓↓␈↓ βWschema and filling it in.}

␈↓ α←␈↓	    Please give the full, formal name for "BLOOD-AGAR"
␈↓ α←␈↓	    ++**␈↓α BLOOD-AGAR␈↓	
␈↓ α←␈↓	    Now please give all synonyms or abbreviations for
␈↓ α←␈↓	    BLOOD-AGAR which you would like the system to accept:
␈↓ α←␈↓	    ++**

␈↓ α←␈↓	Which of the following best describes concepts like BLOOD-AGAR
␈↓ α←␈↓	[Choose the last if no other is appropriate]
␈↓ α←␈↓	     1 - identity of an organism, or
␈↓ α←␈↓	     2 - the site of a culture, or
␈↓ α←␈↓	     3 - None of the above
␈↓ α←␈↓	Choose one
␈↓ α←␈↓	++**␈↓α 3␈↓	

␈↓ α←␈↓↓␈↓ βW{After␈α∞finishing␈α∞with␈α∞the␈α
␈↓	VALUE-SCHEMA␈↓↓,␈α∞the␈α∞system␈α∞is␈α∞again␈α
faced
␈↓ α←␈↓↓␈↓ βWwith␈αa␈αchoice␈αof␈αpaths␈α(see␈αFig.␈α6-3)␈αand␈αmust␈αask␈αfor␈αadvice.␈α Since
␈↓ α←␈↓↓␈↓ βWtwo␈α∀different␈α∪attributes␈α∀can␈α∀have␈α∪values␈α∀in␈α∪common,␈α∀there␈α∀is␈α∪no
␈↓ α←␈↓↓␈↓ βWguarantee␈α⊃that␈α⊃the␈α⊃values␈α⊃of␈α⊂this␈α⊃new␈α⊃attribute␈α⊃are␈α⊃unique.␈α⊂ Also,
␈↓ α←␈↓↓␈↓ βWsince␈α∩the␈α∩phrases␈α∩like␈α∩``the␈α∪site␈α∩of␈α∩a␈α∩culture''␈α∩are␈α∪simply␈α∩strings
␈↓ α←␈↓␈↓168    KNOWLEDGE ACQUISITION II␈↓ 
∪6-10␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βWassociated␈αwith␈αeach␈αschema,␈αthe␈αsystem␈αhas␈αno␈αway␈αof␈αknowing␈αwhat
␈↓ α←␈↓↓␈↓ βWeach␈α∂choice␈α∂represents.␈α∂ The␈α∂expert␈α∂indicates␈α∂that␈α∂the␈α∂concept␈α∂he␈α∂is
␈↓ α←␈↓↓␈↓ βWtalking about is new by choosing the third item.}

␈↓ α←␈↓	Ok, then you'll have to tell me a few things about it...

␈↓ α←␈↓	***************************************************************

␈↓"β␈↓ α←␈↓␈↓ β?Let's␈αtake␈αa␈αmoment␈αout␈αto␈αreview␈αwhat's␈αhappened␈αso␈αfar␈αand␈αto␈αsee
␈↓ α←␈↓where␈αwe␈αare␈α
going␈αfrom␈αhere.␈α
(A␈αrevised␈αversion␈α
of␈αFig.␈α6-3␈α
is␈αreproduced
␈↓ α←␈↓below␈αfor␈αreference.␈α It␈αincludes␈αan␈αindication␈αof␈αthe␈αpath␈αthis␈αexample␈αtakes
␈↓ α←␈↓through the network and omits several other network branches for clarity.)
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂system␈α⊂was␈α⊃able␈α⊂to␈α⊂use␈α⊃its␈α⊂pattern-matching␈α⊂routines␈α⊃to␈α⊂guess
␈↓ α←␈↓from␈α∂the␈α∞dialog␈α∂that␈α∞the␈α∂new␈α∞object␈α∂being␈α∞discussed␈α∂was␈α∞an␈α∂attribute␈α∂of␈α∞a
␈↓ α←␈↓culture.␈α∞ The␈α∂expert␈α∞verified␈α∂this␈α∞guess␈α∞and␈α∂the␈α∞system␈α∂used␈α∞as␈α∂its␈α∞starting
␈↓ α←␈↓point␈α⊂the␈α∂␈↓	CULATTRIB-SCHEMA␈↓,␈α⊂since␈α∂it␈α⊂was␈α∂the␈α⊂schema␈α∂associated␈α⊂with␈α∂the
␈↓ α←␈↓pattern␈α⊃that␈α⊃matched.␈α⊃From␈α⊃there␈α⊃(indicated␈α⊃in␈α⊃Fig.␈α⊃6-9␈α⊃by␈α⊃the␈α⊂asterisk),
␈↓ α←␈↓the␈α
system␈α∞``climbed''␈α
up␈α
one␈α∞level␈α
in␈α
the␈α∞network␈↓
15␈↓␈α
and␈α
started␈α∞back␈α
down.
␈↓ α←␈↓The␈α∂first␈α∂schema␈α∂to␈α∂be␈α∂filled␈α∞out␈α∂is␈α∂the␈α∂␈↓	ATTRIB-SCHEMA␈↓;␈α∂this␈α∂supplied␈α∞the
␈↓ α←␈↓direction␈α∂for␈α∞the␈α∂initial␈α∞part␈α∂of␈α∞the␈α∂dialog.␈α∞Then,␈α∂since␈α∞the␈α∂path␈α∂had␈α∞been
␈↓ α←␈↓marked␈α(during␈αthe␈αascent),␈αit␈α
descended␈αto␈αthe␈α␈↓	CULATTRIB-SCHEMA␈↓␈αand␈α
used
␈↓ α←␈↓that to continue the dialog.

␈↓"β␈↓ α←␈↓	                 KSTRUCT-SCHEMA

␈↓"β␈↓ α←␈↓	   VALUE-SCHEMA                   ATTRIB-SCHEMA
␈↓"β␈↓ α←␈↓	     ??

␈↓"β␈↓ α←␈↓	SITE-       IDENT-
␈↓"β␈↓ α←␈↓	SCHEMA      SCHEMA

␈↓"β␈↓ α←␈↓	  [nutrient-schema]


␈↓"β␈↓ α←␈↓	              PTATTRIB-   INFATTRIB-   CULATTRIB-   ORGATTRIB-
␈↓"β␈↓ α←␈↓	              SCHEMA      SCHEMA       SCHEMA [*]   SCHEMA
␈↓"β␈↓ α←␈↓	                                         ?



␈↓"β␈↓ α←␈↓	                      SVA-SCHEMA  MVA-SCHEMA  TFA-SCHEMA


␈↓"β␈↓ α←␈↓α␈↓ ∧2Fig. 6-9.    Part of the schema hierarchy.    

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[15]␈αSince␈αthe␈α␈↓	KSTRUCT-SCHEMA␈↓␈α
is␈αempty,␈αthere␈αis␈αno␈α
need␈αto␈αgo␈αall␈α
the␈αway
␈↓ α←␈↓to the root.
␈↓ α←␈↓␈↓6-10␈↓ π∂ACQUIRING A NEW ATTRIBUTE    169␈↓

␈↓"β␈↓ α←␈↓␈↓ β?At␈α∂that␈α∂point,␈α∂the␈α∂system␈α∂encountered␈α∂a␈α∂branch␈α∂in␈α∂the␈α∂network␈α∂for
␈↓ α←␈↓which␈α⊂it␈α⊂had␈α⊂no␈α⊃directional␈α⊂information␈α⊂(indicated␈α⊂by␈α⊂the␈α⊃single␈α⊂question
␈↓ α←␈↓mark)␈α⊂and␈α∂hence␈α⊂had␈α⊂to␈α∂ask␈α⊂``␈↓	Which␈α⊂of␈α∂the␈α⊂following␈α⊂best␈α∂describes
␈↓ α←␈↓	concepts␈α→like␈α→NUTRIENT?␈↓'' ␈α→The␈α_process␈α→continued␈α→after␈α→the␈α_expert
␈↓ α←␈↓indicated the correct choice.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α
filling␈αout␈α
this␈α
schema,␈αhowever,␈α
the␈α
system␈αencountered␈α
a␈α
link␈αto
␈↓ α←␈↓another␈αtype␈αof␈αdata␈αstructure.␈αSince␈αeach␈αattribute␈αcarries␈αwith␈αit␈αa␈αlist␈αof␈αits
␈↓ α←␈↓associated␈αvalues,␈αconstructing␈α
a␈αnew␈αattribute␈αleads␈α
to␈αthe␈αacquisition␈αof␈α
new
␈↓ α←␈↓values.  As shown in the trace, this is triggered by encountering

␈↓"β␈↓ α←␈↓	␈↓ βOLEGALVALUES   (KLEENE (1) < VALUE-INST >)   ASKIT

␈↓ α←␈↓in␈αthe␈α␈↓	SVA-SCHEMA␈↓.␈α After␈αlisting␈αall␈αof␈αthe␈αassociated␈αvalues,␈αthe␈αexpert␈αwas
␈↓ α←␈↓asked␈α⊃to␈α∩describe␈α⊃each.␈α∩ The␈α⊃description␈α⊃task␈α∩is␈α⊃set␈α∩up␈α⊃as␈α∩a␈α⊃subproblem
␈↓ α←␈↓(indicated␈α
by␈α
the␈α
dashed␈α
line)␈α
with␈α
the␈α
starting␈α
point␈α
in␈α
the␈α
network␈αgiven␈α
by
␈↓ α←␈↓the schema named in the triple.
␈↓"β␈↓ α←␈↓␈↓ β?Consider␈α
the␈α
description␈α
of␈α∞the␈α
first␈α
value.␈α
 The␈α
system␈α∞started␈α
with
␈↓ α←␈↓the␈α∂␈↓	VALUE-SCHEMA␈↓,␈α∂but␈α∂then␈α∂reached␈α∂a␈α∂branch␈α∂point␈α∂for␈α∂which␈α∂it␈α⊂had␈α∂no
␈↓ α←␈↓information␈α⊂(the␈α⊂double␈α⊂question␈α⊂mark)␈α⊃and␈α⊂again␈α⊂had␈α⊂to␈α⊂ask␈α⊃the␈α⊂expert
␈↓ α←␈↓``␈↓	Which␈α⊂of␈α⊂the␈α⊂following␈α⊂best␈α⊂describes␈α⊂things␈α⊂like␈α⊂BLOOD-AGAR?␈↓'' 
␈↓ α←␈↓This␈αtime,␈αhowever,␈αthe␈αexpert␈α
indicated␈αthat␈αthe␈αobject␈αbeing␈α
acquired␈αwas
␈↓ α←␈↓of␈αa␈α
type␈αnot␈α
yet␈αknown␈αto␈α
the␈αsystem.␈α
 The␈αacquisition␈αof␈α
a␈αnew␈α
schema␈αto
␈↓ α←␈↓describe the new data type is then set up as a sub-subproblem.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
next␈αstep␈α
in␈α
the␈αdialog,␈α
then,␈αwill␈α
be␈α
the␈αdescription␈α
of␈α
the␈αnew
␈↓ α←␈↓``data␈α⊃type''␈α⊃␈↓	NUTRIENT␈↓,␈α⊃accomplished␈α⊂by␈α⊃filling␈α⊃out␈α⊃the␈α⊃␈↓	SCHEMA-SCHEMA␈↓␈α⊂to
␈↓ α←␈↓produce␈αthe␈α␈↓	NUTRIENT-SCHEMA␈↓.␈α This␈αwill␈αbecome␈α
a␈αpart␈αof␈αthe␈αnetwork␈αas␈α
a
␈↓ α←␈↓new␈αbranch␈αbelow␈αthe␈α
␈↓	VALUE-SCHEMA␈↓.␈αNote␈αthat␈αthe␈αnetwork␈α
thus␈α``evolves,''
␈↓ α←␈↓growing␈αlarger␈α
in␈αa␈αrelatively␈α
smooth␈αand␈α
natural␈αway␈αas␈α
the␈αnumber␈αof␈α
data
␈↓ α←␈↓types␈α∂increases.␈α∞ In␈α∂this␈α∞particular␈α∂case,␈α∂almost␈α∞all␈α∂of␈α∞this␈α∂operation␈α∂can␈α∞be
␈↓ α←␈↓carried␈α
out␈α
by␈αthe␈α
system␈α
itself,␈αand␈α
it␈α
is␈α
thus␈αalmost␈α
totally␈α
transparent␈αto␈α
the
␈↓ α←␈↓expert.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
process␈αthen␈α
returns␈αto␈α
the␈αsubproblem␈α
of␈αdescribing␈α
the␈αvalues
␈↓ α←␈↓and␈α
continues␈α
with␈α
␈↓	BLOOD-AGAR␈↓␈α
where␈α
it␈α
left␈α
off.␈α
That␈α
is,␈α
after␈α
describing␈α
the
␈↓ α←␈↓new␈α∂data␈α∞type␈α∂(sprouting␈α∞the␈α∂new␈α∞branch␈α∂of␈α∞the␈α∂network),␈α∞the␈α∂process␈α∞will
␈↓ α←␈↓continue␈α∂down␈α∂into␈α∞that␈α∂branch␈α∂and␈α∞will␈α∂request␈α∂a␈α∞description␈α∂of␈α∂the␈α∞first
␈↓ α←␈↓instance of the new data type, using the new schema as a guide.
␈↓"β␈↓ α←␈↓␈↓ β?After␈α⊃all␈α⊃the␈α⊂values␈α⊃have␈α⊃been␈α⊃described,␈α⊂the␈α⊃dialog␈α⊃goes␈α⊃back␈α⊂to
␈↓ α←␈↓finish␈α
up␈α
acquisition␈α
of␈α
the␈α
new␈α
attribute.␈α
 The␈α
entire␈α
sequence␈α
of␈α∞topics␈α
is
␈↓ α←␈↓shown below.
␈↓ α←␈↓␈↓170    KNOWLEDGE ACQUISITION II␈↓ 
∪6-10␈↓


␈↓"β␈↓ α←␈↓	New rule ␈↓↓acquisition requires a␈↓	
␈↓"β␈↓ α←␈↓	    new attribute ␈↓↓ which has its associated␈↓	
␈↓"β␈↓ α←␈↓	          new values.  ␈↓↓But these are a␈↓	
␈↓"β␈↓ α←␈↓	                new data type ␈↓↓which means adding a new schema to the␈↓	
␈↓"β␈↓ α←␈↓	                              ␈↓↓network.␈↓	
␈↓"β␈↓ α←␈↓	          New values ␈↓↓then instantiate the new schema, and finally the␈↓	
␈↓"β␈↓ α←␈↓	    new attribute ␈↓↓is finished up, so the system returns to the␈↓	
␈↓"β␈↓ α←␈↓	new rule.

␈↓"β␈↓ α←␈↓␈↓ β?Let's␈α∂continue␈α⊂with␈α∂the␈α⊂trace,␈α∂picking␈α∂it␈α⊂up␈α∂at␈α⊂the␈α∂point␈α⊂where␈α∂the
␈↓ α←␈↓system␈α∩requested␈α∩help␈α∩in␈α∩choosing␈α∩the␈α∩path␈α∩to␈α∩follow␈α∩below␈α∪the␈α∩␈↓	VALUE-
␈↓ α←␈↓	SCHEMA␈↓.

␈↓ α←␈↓	***************************************************************

␈↓ α←␈↓	Which of the following best describes things like BLOOD-AGAR
␈↓ α←␈↓	[Choose the last if no other is appropriate]
␈↓ α←␈↓	     1 - identity of an organism, or
␈↓ α←␈↓	     2 - the site of a culture, or
␈↓ α←␈↓	     3 - None of the above
␈↓ α←␈↓	Choose one
␈↓ α←␈↓	++** ␈↓α3␈↓	

␈↓ α←␈↓	Ok, then you'll have to tell me a few things about it...

␈↓ α←␈↓↓␈↓ βW{␈↓¬TEIRESIAS␈↓↓␈α
retrieves␈α
the␈α
␈↓	SCHEMA-SCHEMA␈↓↓␈α
to␈α
direct␈α
the␈α
acquisition␈αof␈α
the
␈↓ α←␈↓↓␈↓ βWnew data type.}

␈↓ α←␈↓	    Please give a phrase (like the ones above) to describe this
␈↓ α←␈↓	    new type of value.
␈↓ α←␈↓	    ++** ␈↓αTHE NUTRIENT OF A CULTURE MEDIUM␈↓	


␈↓ α←␈↓	    Ok, finished describing the new type of value; now back to
␈↓ α←␈↓	    BLOOD-AGAR

␈↓ α←␈↓↓␈↓ βW{In␈αthis␈αcase,␈αmost␈αof␈αthe␈αtask␈αof␈αconstructing␈αthe␈αnew␈αschema␈αcan␈αbe
␈↓ α←␈↓↓␈↓ βWhandled␈α
by␈α∞the␈α
system␈α∞itself,␈α
so␈α
the␈α∞dialog␈α
returns␈α∞to␈α
the␈α∞problem␈α
of
␈↓ α←␈↓↓␈↓ βWdescribing the new instances of the (new) data type.}

␈↓ α←␈↓	  Finished with BLOOD-AGAR; now THAYER-MARTIN

␈↓ α←␈↓↓␈↓ βW{And␈α∞there's␈α∞nothing␈α∞else␈α∞to␈α∂do␈α∞in␈α∞describing␈α∞the␈α∞first␈α∞value,␈α∂so␈α∞the
␈↓ α←␈↓↓␈↓ βWsystem continues with the second.}

␈↓ α←␈↓	    Please give the full, formal name for "THAYER-MARTIN"
␈↓ α←␈↓	    ++** ␈↓αTHAYER-MARTIN␈↓	
␈↓"β␈↓ α←␈↓␈↓6-10␈↓ π∂ACQUIRING A NEW ATTRIBUTE    171␈↓

␈↓"β␈↓ α←␈↓	    Now please give all synonyms or abbreviations for
␈↓ α←␈↓	    THAYER-MARTIN which you would like the system to accept:
␈↓ α←␈↓	    ++** ␈↓αCHOCOLATE-AGAR␈↓	
␈↓ α←␈↓	    ++** ␈↓αTM␈↓	
␈↓ α←␈↓	    ++**

␈↓ α←␈↓	    Finished with THAYER-MARTIN
␈↓ α←␈↓	  OK done with the values for NUTRIENT now...

␈↓ α←␈↓↓␈↓ βW{The␈α⊃second␈α⊃instance␈α⊃is␈α⊃likewise␈α⊂described␈α⊃in␈α⊃short␈α⊃order,␈α⊃and␈α⊂the
␈↓ α←␈↓↓␈↓ βWdialog␈α∞returns␈α
to␈α∞the␈α∞original␈α
problem--describing␈α∞the␈α∞new␈α
attribute.
␈↓ α←␈↓↓␈↓ βWThere is just one more part to be acquired.}

␈↓ α←␈↓	    Please give a phrase which can be used to ask about the
␈↓ α←␈↓	    value of NUTRIENT
␈↓ α←␈↓	    [Type as many answers as are appropriate, then an empty
␈↓ α←␈↓	     line.]
␈↓ α←␈↓	    ++** ␈↓αWHAT IS THE NUTRIENT OF THE MEDIUM FOR *␈↓	
␈↓ α←␈↓	    ++**

␈↓ α←␈↓	Ok, finished with describing the new attribute now.
␈↓ α←␈↓	Returning to parsing the new rule.

␈↓ α←␈↓	This may take a few moments...


␈↓ α←␈↓	This is my understanding of your rule:
␈↓ α←␈↓	RULE383
␈↓ α←␈↓	-------
␈↓ α←␈↓	  If:  1) The gramstain of the organism is gramnegative, and
␈↓ α←␈↓	       2) The morphology of the organism is rod, and
␈↓ α←␈↓	       3) The patient is a compromised host, and
␈↓ α←␈↓	       4) The nutrient of the medium of the culture is
␈↓ α←␈↓	          blood-agar
␈↓ α←␈↓	  Then:  There is strongly suggestive evidence (.8) that the
␈↓ α←␈↓	         category of the organism is enterobacteriaceae


␈↓ α←␈↓	Okay? (Yes or No)
␈↓ α←␈↓	++** ␈↓αY␈↓	

␈↓ α←␈↓↓␈↓ βW{Finally, we have the new rule, with its new attribute and value.}

␈↓"β␈↓ α←␈↓␈↓α6-10-1    Comments on the trace␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αcreation␈αof␈αthe␈αnew␈αattribute␈αis␈αan␈αextensive␈αoperation␈αinvolving
␈↓ α←␈↓several␈α⊂different␈α⊂data␈α∂types.␈α⊂ By␈α⊂structuring␈α∂it␈α⊂properly,␈α⊂however,␈α⊂the␈α∂task
␈↓ α←␈↓becomes␈α⊗comprehensible.␈α⊗ There␈α⊗are␈α⊗several␈α⊗sources␈α⊗of␈α⊗this␈α⊗structuring
␈↓ α←␈↓information.␈α∂ First,␈α∂the␈α∂schemata␈α∂take␈α∂advantage␈α∂of␈α∂the␈α∂decomposability␈α∞of
␈↓ α←␈↓individual␈α∩data␈α∩types␈α∩to␈α∩present␈α∩a␈α∩series␈α∩of␈α∩straightforward,␈α⊃independent
␈↓"β␈↓ α←␈↓␈↓172    KNOWLEDGE ACQUISITION II␈↓ 
∪6-10␈↓

␈↓"β␈↓ α←␈↓questions.␈α∩ Next,␈α∩the␈α∩schema␈α∩network␈α∩relies␈α∩on␈α∩the␈α∪fundamentally␈α∩simple
␈↓ α←␈↓organization␈α∩of␈α∪the␈α∩data␈α∪types␈α∩to␈α∩provide␈α∪a␈α∩comprehensible␈α∪sequence␈α∩of
␈↓ α←␈↓topics.␈α The␈αslots␈αand␈αslotexperts,␈αin␈αturn,␈αmake␈αit␈αpossible␈αto␈αrepresent␈αmany
␈↓ α←␈↓conventions␈α⊃of␈α⊂the␈α⊃data␈α⊂types␈α⊃in␈α⊂ways␈α⊃that␈α⊂permit␈α⊃the␈α⊂system␈α⊃to␈α⊂perform
␈↓ α←␈↓many␈α∪of␈α∩the␈α∪routine␈α∩tasks,␈α∪considerably␈α∩simplifying␈α∪the␈α∪entire␈α∩operation.
␈↓ α←␈↓Finally,␈α∂the␈α⊂correspondence␈α∂between␈α⊂data␈α∂types␈α∂and␈α⊂objects␈α∂in␈α⊂the␈α∂domain
␈↓ α←␈↓makes␈α⊃it␈α⊃possible␈α⊃to␈α⊃present␈α⊃a␈α⊃dialog␈α⊃that␈α⊃appears␈α⊃comprehensible␈α⊃to␈α⊂the
␈↓ α←␈↓expert,␈α≠yet␈α≠which␈α≠deals␈α≠effectively␈α≠with␈α≠questions␈α≠of␈α≠data␈α≠structure
␈↓ α←␈↓manipulation.␈α The␈α
result␈αof␈α
all␈αthis␈α
is␈αthe␈α
construction␈αof␈α
some␈αcomplex␈α
data
␈↓ α←␈↓structures␈α⊂with␈α⊂numerous␈α⊃internal␈α⊂conventions␈α⊂and␈α⊂interrelationships,␈α⊃in␈α⊂a
␈↓ α←␈↓fashion that makes it a reasonable task for the expert.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂growth␈α∂of␈α⊂the␈α⊂schema␈α∂network␈α⊂to␈α⊂encompass␈α∂a␈α⊂new␈α⊂data␈α∂type
␈↓ α←␈↓demonstrates␈α⊂the␈α∂degree␈α⊂of␈α⊂flexibility␈α∂in␈α⊂the␈α∂system.␈α⊂ The␈α⊂flexibility␈α∂arises
␈↓ α←␈↓from␈α⊗the␈α⊗use␈α∃of␈α⊗the␈α⊗schemata␈α∃as␈α⊗a␈α⊗language␈α∃and␈α⊗framework␈α⊗for␈α∃the
␈↓ α←␈↓specification␈αof␈αrepresentations.␈α
 Knowledge␈αabout␈αany␈αspecific␈α
representation
␈↓ α←␈↓is␈α⊃contained␈α⊃entirely␈α⊃in␈α⊃the␈α⊃``statements''␈α⊃of␈α⊃that␈α⊃language,␈α⊃rather␈α∩than␈α⊃in
␈↓ α←␈↓special␈α⊃purpose␈α⊂code.␈α⊃ This␈α⊃provides␈α⊂a␈α⊃greater␈α⊂range␈α⊃of␈α⊃applicability␈α⊂and
␈↓ α←␈↓flexibility␈α∃than␈α∃would␈α∃be␈α∀possible␈α∃if␈α∃separate,␈α∃hand-tailored␈α∀acquisition
␈↓ α←␈↓routines were written for each different data type.
␈↓"β␈↓ α←␈↓␈↓ β?Additional␈α∞flexibility␈α
arises␈α∞from␈α∞the␈α
inherently␈α∞extensible␈α∞nature␈α
of
␈↓ α←␈↓the␈αschema␈αnetwork.␈α As␈αwith␈α
all␈αgeneralization␈αhierarchies,␈αit␈αis␈α
a␈αrelatively
␈↓ α←␈↓simple␈αoperation␈αto␈αadd␈αnew␈αbranches␈αat␈αany␈αlevel␈αin␈αthe␈αnetwork.␈α Since␈αthe
␈↓ α←␈↓representation␈α∃language␈α∃interpreter␈α∃``reads''␈α∀the␈α∃network␈α∃to␈α∃structure␈α∀the
␈↓ α←␈↓dialog,␈α
the␈α
addition␈α
will␈α
be␈αreflected␈α
in␈α
future␈α
acquisition␈α
sessions.␈α
 That␈αis,
␈↓ α←␈↓the␈α
next␈αtime␈α
␈↓¬TEIRESIAS␈↓␈α
reaches␈αthe␈α
␈↓	VALUE-SCHEMA␈↓␈αnode␈α
and␈α
requests␈αadvice
␈↓ α←␈↓about␈α
which␈α
way␈αto␈α
go,␈α
it␈αwill␈α
present␈α
the␈α
expert␈αwith␈α
four␈α
options: ␈αthe␈α
three
␈↓ α←␈↓shown in the previous trace plus the new one of culture medium nutrient.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂trace␈α∞also␈α∂demonstrates␈α∂that␈α∞the␈α∂description␈α∞of␈α∂the␈α∂structure␈α∞of
␈↓ α←␈↓one␈α∀data␈α∀type␈α∪may␈α∀mention␈α∀another␈α∀(as␈α∪the␈α∀description␈α∀of␈α∀an␈α∪␈↓↓attribute␈↓
␈↓ α←␈↓mentions␈α∞␈↓↓values␈↓).␈α∞ In␈α∞the␈α∞acquisition␈α∞process␈α∞this␈α∞gets␈α∞translated␈α∞into␈α∞a␈α
new
␈↓ α←␈↓direction␈α⊃for␈α⊃the␈α⊂dialog,␈α⊃as␈α⊃one␈α⊃topic␈α⊂(describing␈α⊃the␈α⊃new␈α⊃attribute)␈α⊂leads
␈↓ α←␈↓naturally␈α→into␈α→another␈α~(describing␈α→its␈α→associated␈α→values).␈α~ These␈α→``new
␈↓ α←␈↓directions''␈αare␈αcurrently␈αfollowed␈αas␈αthey␈αarise␈α(i.e.,␈αthe␈αsearch␈αis␈αdepth-first).
␈↓ α←␈↓This␈α∞can␈α∞prove␈α∞to␈α∞be␈α∞a␈α∞distraction␈α∞at␈α∞times,␈α∞since␈α∞the␈α∞dialog␈α∞goes␈α∞off␈α∞on␈α∞a
␈↓ α←␈↓subtopic␈α
and␈αlater␈α
returns␈α
to␈αthe␈α
main␈αtopic␈α
to␈α
finish␈αup.␈α
 This␈α
could␈αeasily
␈↓ α←␈↓be␈α∩changed␈α∩to␈α∩a␈α∩modified␈α⊃breadth-first␈α∩search,␈α∩which␈α∩would␈α∩result␈α∩in␈α⊃a
␈↓ α←␈↓dialog␈α⊗that␈α⊗exhausted␈α⊗each␈α↔topic␈α⊗(each␈α⊗data␈α⊗structure)␈α⊗in␈α↔turn␈α⊗before
␈↓ α←␈↓beginning another.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∪final␈α∪comment␈α∪concerns␈α∪the␈α∪simplicity␈α∪of␈α∪acquiring␈α∪the␈α∪new
␈↓ α←␈↓schema␈α⊂that␈α⊂describes␈α⊂the␈α⊂values␈α⊂of␈α⊂the␈α⊂new␈α⊂attribute.␈α⊂ There␈α⊃are␈α⊂several
␈↓ α←␈↓reasons␈α∂why␈α⊂the␈α∂operation␈α∂is␈α⊂in␈α∂this␈α∂case␈α⊂almost␈α∂totally␈α∂transparent␈α⊂to␈α∂the
␈↓ α←␈↓expert␈αwhen␈αin␈αgeneral␈αit␈αis␈αa␈αmuch␈αmore␈αcomplex␈αoperation.␈α It␈αis␈αin␈αpart␈αa
␈↓ α←␈↓fortuitous␈α→side␈α_effect␈α→of␈α→the␈α_conventions␈α→used␈α→in␈α_the␈α→current␈α→set␈α_of
␈↓ α←␈↓representations.␈α≡ Most␈α≡of␈α≡the␈α≡important␈α≡conventions␈α∨concerning␈α≡the
␈↓ α←␈↓representation␈αof␈αa␈αvalue␈αare␈α
common␈αto␈αall␈αvalues␈αand␈αhence␈α
are␈αexpressed
␈↓ α←␈↓␈↓6-10␈↓ π∂ACQUIRING A NEW ATTRIBUTE    173␈↓

␈↓"β␈↓ α←␈↓in␈α∞the␈α∞network␈α∞at␈α∞the␈α∂level␈α∞of␈α∞the␈α∞␈↓	VALUE-SCHEMA␈↓.␈α∞ There␈α∞is␈α∂thus␈α∞relatively
␈↓ α←␈↓little more that the schemata at lower levels have to add.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∂transparency␈α⊂also␈α∂results␈α∂from␈α⊂the␈α∂assumption␈α∂that␈α⊂the␈α∂expert
␈↓ α←␈↓will␈α⊗not␈α⊗be␈α⊗expected␈α⊗to␈α⊗make␈α∃changes␈α⊗in␈α⊗the␈α⊗basic␈α⊗formalism␈α⊗of␈α∃the
␈↓ α←␈↓performance␈α
program.␈α
 In␈α
line␈α
with␈α
this␈α
assumption,␈α
when␈α
dealing␈α∞with␈α
the
␈↓ α←␈↓expert,␈αthe␈αschema␈αinterpreter␈αdoes␈αnot␈αrequest␈αtwo␈αtypes␈αof␈αinformation␈αthat
␈↓ α←␈↓would␈α
normally␈α
be␈α
part␈αof␈α
the␈α
description␈α
of␈αa␈α
new␈α
data␈α
type: ␈αsubstructure
␈↓ α←␈↓and␈α⊂interrelationships.␈α⊂ Wherever␈α⊂a␈α⊂new␈α⊂schema␈α⊂is␈α⊂added␈α⊂to␈α⊂the␈α∂network,
␈↓ α←␈↓there␈α⊃is␈α⊃the␈α⊃possibility␈α⊃that␈α⊃the␈α⊃data␈α⊃structure␈α⊃it␈α⊃describes␈α⊃has␈α⊃additional
␈↓ α←␈↓substructure␈αand␈αinterrelationships␈αbeyond␈α
those␈αdescribed␈αby␈αits␈αancestors␈α
in
␈↓ α←␈↓the␈α∞network.␈α∞ To␈α∞be␈α∞complete,␈α∂the␈α∞system␈α∞should␈α∞naturally␈α∞ask␈α∂about␈α∞them.
␈↓ α←␈↓But␈α↔notice␈α⊗that␈α↔the␈α↔answer␈α⊗to␈α↔either␈α↔the␈α⊗question␈α↔of␈α↔substructure␈α⊗or
␈↓ α←␈↓interrelationship␈α
requires␈αa␈α
knowledge␈αof,␈α
and␈αimplies␈α
potential␈αalterations␈α
to,
␈↓ α←␈↓the␈αunderlying␈αformalism␈αof␈αthe␈αperformance␈αprogram.␈α Any␈α
substructure␈αin
␈↓ α←␈↓a␈αnew␈αdata␈αtype␈αwould␈αhave␈αto␈αbe␈αreferenced␈αsomewhere␈αin␈αthe␈αperformance
␈↓ α←␈↓program␈α∞if␈α∞it␈α∞is␈α∞to␈α∞be␈α∞of␈α∞use,␈α∞implying␈α∞that␈α∞the␈α∞performance␈α∞program␈α
code
␈↓ α←␈↓would␈α⊃have␈α⊃to␈α⊃be␈α⊃changed.␈α⊃ The␈α⊃ability␈α⊃to␈α⊃specify␈α⊃new␈α⊂interrelationships
␈↓ α←␈↓between␈α
data␈α
types␈α
implies␈α
an␈α∞understanding␈α
of␈α
the␈α
data␈α
types␈α∞that␈α
already
␈↓ α←␈↓exist.␈α
 Since␈α
both␈αof␈α
these␈α
clearly␈α
require␈αan␈α
understanding␈α
of␈α
elements␈αof␈α
the
␈↓ α←␈↓system␈αthat␈αwould␈αbe␈αalien␈αto␈αthe␈αexpert,␈αthey␈αare␈αomitted.␈α (We␈αwill␈αsee␈αlater
␈↓ α←␈↓that they are asked under other circumstances.)
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∪approach␈α∀allows␈α∪the␈α∀expert␈α∪to␈α∪teach␈α∀the␈α∪system␈α∀about␈α∪new
␈↓ α←␈↓attributes␈αand␈αvalues␈αwithout␈αgetting␈αinvolved␈αin␈αprogramming␈αdetails.␈α The
␈↓ α←␈↓price␈αis␈αa␈αsmall␈α
possibility␈αthat␈αhe␈αmay␈α
compromise␈αthe␈αintegrity␈αof␈α
the␈αdata
␈↓ α←␈↓base, if the new data type in fact should be related to some existing structure.
␈↓"β␈↓ α←␈↓␈↓ β?There␈αis␈α
really␈αa␈α
more␈αfundamental␈α
problem␈αhere:␈α
The␈αcurrent␈α
design
␈↓ α←␈↓of␈α␈↓¬MYCIN␈↓'s␈αrepresentations␈αmakes␈αeach␈αkind␈αof␈α␈↓	VALUE␈↓␈αits␈αown␈αdata␈αtype.␈α This
␈↓ α←␈↓is␈αwhat␈αmakes␈αit␈αnecessary␈αto␈αacquire␈αa␈αnew␈αschema␈αand␈αpushes␈αthe␈αtask␈αinto
␈↓ α←␈↓the␈α∞realm␈α∞of␈α∞changes␈α∞to␈α∞the␈α∞performance␈α∞program.␈α∞ With␈α∞some␈α∂redesign␈α∞of
␈↓ α←␈↓the␈αdata␈α
structures␈αinvolved,␈αit␈α
would␈αbe␈α
possible␈αto␈αhave␈α
just␈αa␈α
single␈αkind
␈↓ α←␈↓of␈α␈↓	VALUE␈↓␈αdata␈αtype,␈αand␈αavoid␈αall␈αthis.␈α But␈αas␈αindicated,␈αit␈αwas␈αnecessary␈αto
␈↓ α←␈↓work␈α
within␈α
the␈αexisting␈α
representations␈α
in␈α
␈↓¬MYCIN␈↓␈αand␈α
still␈α
make␈α
it␈αpossible␈α
for
␈↓ α←␈↓the expert to educate the system.

␈↓"β␈↓ α←␈↓␈↓α6-11    KNOWLEDGE ABOUT KNOWLEDGE ABOUT
␈↓ α←␈↓α␈↓ β3REPRESENTATIONS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓␈α_was␈α_designed␈α_to␈α_make␈α_possible␈α_interactive␈α_transfer␈α↔of
␈↓ α←␈↓expertise.␈αAs␈αwe␈αhave␈α
seen,␈αone␈αkind␈αof␈α
expertise␈αit␈αcan␈αtransfer␈α
is␈αdomain-
␈↓ α←␈↓specific␈α
information,␈α
the␈α
kind␈α
supplied␈α
by␈α
an␈α
expert␈α
to␈α
improve␈α
the␈α
operation
␈↓ α←␈↓of␈αa␈αperformance␈αprogram.␈α But␈αrecall␈αthat␈αhigh␈αperformance␈αon␈αthe␈α
transfer
␈↓ α←␈↓of␈α∩expertise␈α∩task␈α⊃required␈α∩a␈α∩store␈α⊃of␈α∩knowledge␈α∩about␈α∩representations.␈α⊃If
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓␈α⊗is␈α⊗designed␈α⊗to␈α↔make␈α⊗possible␈α⊗interactive␈α⊗transfer␈α↔of␈α⊗expertise
␈↓ α←␈↓independent␈α∪of␈α∪domain,␈α∩why␈α∪not␈α∪apply␈α∩it␈α∪to␈α∪the␈α∩task␈α∪of␈α∪acquiring␈α∩and
␈↓ α←␈↓maintaining␈αthe␈αrequisite␈αbase␈αof␈αknowledge␈αabout␈αrepresentations?␈α That␈αis,
␈↓ α←␈↓why␈α_not␈α_push␈α→this␈α_back␈α_a␈α_level␈α→and␈α_consider␈α_the␈α→knowledge␈α_about
␈↓ α←␈↓␈↓174    KNOWLEDGE ACQUISITION II␈↓ 
∪6-11␈↓

␈↓"β␈↓ α←␈↓representations␈αas␈αa␈αcandidate␈αfor␈αinteractive␈αtransfer␈αof␈αexpertise? ␈αThis␈αhas
␈↓ α←␈↓been done and involves using ␈↓¬TEIRESIAS␈↓ in two phases (Fig. 6-10).
␈↓"β␈↓ α←␈↓␈↓ β?As␈α∪we␈α∪have␈α∪seen,␈α∪the␈α∀domain␈α∪expert␈α∪uses␈α∪␈↓¬TEIRESIAS␈↓␈α∪to␈α∀teach␈α∪the
␈↓ α←␈↓performance␈α∞program␈α∞about␈α∂the␈α∞domain␈α∞of␈α∞application.␈α∂ High␈α∞performance
␈↓ α←␈↓on␈αthis␈αtask␈αis␈αmade␈αpossible␈αby␈αthe␈αbase␈αof␈αknowledge␈αabout␈αrepresentations
␈↓ α←␈↓provided␈αby␈αthe␈αschemata.␈α But␈αthe␈αsystem␈αarchitect␈αcan␈αalso␈αuse␈α␈↓¬TEIRESIAS␈↓␈αto
␈↓ α←␈↓teach␈α∂about␈α∂a␈α∂particular␈α⊂set␈α∂of␈α∂representations.␈α∂ High␈α∂performance␈α⊂on␈α∂this
␈↓ α←␈↓task␈α⊃is␈α⊃made␈α⊃possible␈α⊃by␈α⊂the␈α⊃␈↓↓schema-schema␈↓,␈α⊃a␈α⊃base␈α⊃of␈α⊃``knowledge␈α⊂about
␈↓ α←␈↓knowledge␈α∩about␈α∩representations,''␈α∩which␈α∩is␈α∩used␈α∩to␈α∩guide␈α∩the␈α∪process␈α∩of
␈↓ α←␈↓describing␈α⊗a␈α⊗new␈α⊗representation.␈α⊗ It␈α⊗is,␈α⊗in␈α⊗effect,␈α⊗a␈α⊗set␈α↔of␈α⊗instructions
␈↓ α←␈↓describing␈αhow␈α
to␈αspecify␈αa␈α
representation.␈α Since␈αthe␈α
instructions␈αare␈α
in␈αthe
␈↓ α←␈↓same␈αformat␈αas␈α
those␈αin␈αan␈α
ordinary␈αschema,␈αthe␈α
process␈αof␈αfollowing␈αthem␈α
is
␈↓ α←␈↓identical.␈α As␈αa␈αresult,␈αwe␈αneed␈αonly␈αa␈αsingle␈α``schema␈α
interpretation''␈αprocess.
␈↓ α←␈↓Teaching␈α≠about␈α≠a␈α≠representation␈α~(acquiring␈α≠a␈α≠new␈α≠schema)␈α≠is␈α~thus
␈↓ α←␈↓computationally␈α⊃identical␈α⊃to␈α⊃teaching␈α∩about␈α⊃the␈α⊃domain␈α⊃(acquiring␈α∩a␈α⊃new
␈↓ α←␈↓instance␈α∂of␈α∞a␈α∂schema);␈α∞indeed,␈α∂both␈α∂teaching␈α∞tasks␈α∂shown␈α∞in␈α∂Fig.␈α∂6-10␈α∞are
␈↓ α←␈↓done with a single body of code.


␈↓"␈↓ α←␈↓∧            teaching about        teaching about the
␈↓"␈↓ α←␈↓∧           a representation      domain of application
␈↓"␈↓ α←␈↓∧⊂ααααααααααααααα⊃      ⊂ααααααααααααααα⊃       ⊂αααααααααααααα⊃
␈↓"␈↓ α←␈↓∧~knowledge of   ~      ~knowledge of   ~       ~              ~
␈↓"␈↓ α←␈↓∧~representation-~      ~primitives for ~       ~ object-level ~
␈↓"␈↓ α←␈↓∧~independent    ~= = @ ~a specific     ~ = = @ ~knowledge base~
␈↓"␈↓ α←␈↓∧~primitives     ~      ~representation ~       ~              ~
␈↓"␈↓ α←␈↓∧%ααααααααααααααα$      %ααααααααααααααα$       %αααααααααααααα$

␈↓"␈↓ α←␈↓∧   SYSTEM 2                SYSTEM 1                SYSTEM 0


␈↓"␈↓ α←␈↓α␈↓ β<Fig. 6-10.    The two applications of schema instantiation.    

␈↓"β␈↓ α←␈↓␈↓ β?Earlier␈α⊃sections␈α⊃of␈α⊃this␈α⊂chapter␈α⊃displayed␈α⊃three␈α⊃examples␈α⊃from␈α⊂the
␈↓ α←␈↓process␈α
of␈α
teaching␈α
about␈α
the␈αdomain␈α
and␈α
demonstrated␈α
how␈α
each␈α
could␈αbe
␈↓ α←␈↓understood␈α∪in␈α∪terms␈α∪of␈α∪filling␈α∪out␈α∪one␈α∪or␈α∪more␈α∪schemata.␈α∀ This␈α∪section
␈↓ α←␈↓explores␈α
an␈α∞example␈α
of␈α
the␈α∞first␈α
process--teaching␈α
about␈α∞a␈α
representation--
␈↓ α←␈↓and views it in terms of ␈↓↓augmenting the schema network␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∪view␈α∀is␈α∪useful␈α∪both␈α∀computationally␈α∪and␈α∀conceptually.␈α∪ The
␈↓ α←␈↓computational␈α
task␈α
is␈α
simplified␈α
because␈α
much␈α
is␈α
accomplished␈α
by␈α∞adding␈α
a
␈↓ α←␈↓single␈αbranch␈αto␈αthe␈αschema␈αnetwork,␈αwhich␈αis␈αan␈αinformation-rich␈αstructure.
␈↓ α←␈↓The␈α∀new␈α∃schema␈α∀will␈α∃inherit␈α∀all␈α∃of␈α∀the␈α∃information␈α∀represented␈α∃in␈α∀its
␈↓ α←␈↓ancestors␈αin␈α
the␈αnetwork␈α
and␈αhence␈αneed␈α
not␈αreplicate␈α
it.␈α The␈α
task␈αbecomes
␈↓ α←␈↓easier␈α~conceptually␈α~since␈α→the␈α~network␈α~offers␈α→a␈α~useful␈α~framework␈α→for
␈↓ α←␈↓organizing and understanding all the different representations in a program.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α
approach␈α
assumes,␈α∞of␈α
course,␈α
that␈α
the␈α∞different␈α
data␈α
types␈α∞in␈α
a
␈↓ α←␈↓program␈αcan,␈αin␈αfact,␈αbe␈αorganized␈αinto␈αa␈αgeneralization␈αhierarchy.␈α If␈αthis␈αis
␈↓"β␈↓ α←␈↓␈↓6-11␈↓ ∧"KNOWLEDGE ABOUT KNOWLEDGE ABOUT REPRESENTATIONS    175␈↓

␈↓"β␈↓ α←␈↓true,␈α⊂the␈α⊃hierarchy␈α⊂can␈α⊃be␈α⊂used␈α⊂to␈α⊃provide␈α⊂another␈α⊃tool␈α⊂for␈α⊃dealing␈α⊂with
␈↓ α←␈↓complexity,␈α⊃providing␈α⊃a␈α⊃useful␈α⊃organizational␈α⊃overview␈α⊃when␈α⊃there␈α∩are␈α⊃a
␈↓ α←␈↓number of related data types.
␈↓"β␈↓ α←␈↓␈↓ β?Building␈α⊃the␈α⊃schema␈α⊃network␈α∩also␈α⊃provides␈α⊃one␈α⊃useful␈α⊃test␈α∩of␈α⊃the
␈↓ α←␈↓generality␈α
of␈α
this␈α
part␈α
of␈α
the␈α
system.␈α
 If␈α
the␈α
techniques␈α
used␈α
are␈α
sufficiently
␈↓ α←␈↓general,␈α
it␈α
should␈αbe␈α
possible␈α
to␈αgrow␈α
the␈α
entire␈αnetwork␈α
from␈α
a␈αfoundation
␈↓ α←␈↓that␈α
is␈α∞not␈α
specific␈α
to␈α∞any␈α
particular␈α
representation.␈α∞ This␈α
was␈α
made␈α∞one␈α
of
␈↓ α←␈↓the design criteria for the system, and has provided useful guidance.
␈↓"β␈↓ α←␈↓␈↓ β?We␈αwill␈αsee␈αthat␈αthe␈αschemata␈αhave␈αbeen␈αapplied␈αto␈αa␈αvariety␈αof␈αdata
␈↓ α←␈↓structures.␈α∞This␈α∞will␈α∞help␈α∞make␈α∞plausible␈α
the␈α∞claim␈α∞that␈α∞they␈α∞form␈α∞a␈α
useful
␈↓ α←␈↓tool␈αfor␈αattacking␈α
one␈αcentral␈αproblem␈αfaced␈α
in␈αbuilding␈αknowledge␈αbases:␈α
the
␈↓ α←␈↓construction and maintenance of large collections of varied data structures.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈αthis␈αstage␈αof␈αknowledge␈αbase␈αconstruction␈αdeals␈αwith␈αthe␈αprocess
␈↓ α←␈↓of␈α
describing␈αrepresentations,␈α
we␈αdo␈α
not␈αexpect␈α
that␈αit␈α
would␈αbe␈α
accomplished
␈↓ α←␈↓by␈α
the␈αexpert␈α
from␈α
the␈αapplication␈α
domain.␈α
 The␈αtask␈α
requires␈α
a␈αknowledge
␈↓ α←␈↓of␈α⊃programming␈α⊃and␈α⊃may␈α⊃require␈α⊃changes␈α⊃to␈α⊃the␈α⊃basic␈α⊃formalism␈α∩of␈α⊃the
␈↓ α←␈↓performance␈α⊃program.␈α⊃ In␈α⊂this␈α⊃use,␈α⊃then,␈α⊂the␈α⊃schemata␈α⊃are␈α⊃more␈α⊂properly
␈↓ α←␈↓viewed␈α∞as␈α∞a␈α∂``programmer's␈α∞assistant''␈α∞tool,␈α∞to␈α∂be␈α∞used␈α∞by␈α∞someone␈α∂with␈α∞the
␈↓ α←␈↓appropriate␈α∂background.␈α∂ The␈α∞language␈α∂of␈α∂the␈α∞next␈α∂dialog␈α∂will␈α∂reflect␈α∞this
␈↓ α←␈↓new␈α≡orientation,␈α≡since␈α≡it␈α≡assumes␈α≡a␈α≡familiarity␈α≡with␈α≡both␈α≡general
␈↓ α←␈↓programming issues and the language of the schemata.

␈↓"β␈↓ α←␈↓␈↓α6-11-1    The SCHEMA-SCHEMA␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈α∞noted␈α∞earlier,␈α∞in␈α∞our␈α∞framework␈α∞the␈α∞process␈α∞of␈α∞describing␈α∞a␈α∞new
␈↓ α←␈↓representation␈α∂can␈α∂be␈α∂made␈α⊂computationally␈α∂identical␈α∂to␈α∂that␈α⊂of␈α∂describing
␈↓ α←␈↓new␈α∞instances␈α∞of␈α∞a␈α∞representation.␈α∞This␈α∞uniformity␈α∞is␈α∞made␈α∞possible␈α∂by␈α∞the
␈↓ α←␈↓schema-schema, shown below.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
schema-schema,␈α
along␈α
with␈α
some␈α
associated␈α
structures,␈α
provides␈α
a
␈↓ α←␈↓foundation␈α⊂of␈α⊂representation-independent␈α⊂knowledge␈α⊂that␈α⊂can␈α⊂be␈α⊃used␈α⊂for
␈↓ α←␈↓constructing␈α∀an␈α∃entire␈α∀knowledge␈α∃base.␈α∀ The␈α∀nature␈α∃and␈α∀extent␈α∃of␈α∀this
␈↓ α←␈↓knowledge␈αis␈αoutlined␈α
below,␈αto␈αcharacterize␈α
the␈αassumptions␈αbehind␈α
the␈αuse
␈↓ α←␈↓of␈α
the␈α
schema-schema␈αand,␈α
hence,␈α
the␈α
range␈αof␈α
representations␈α
for␈α
which␈αit␈α
is
␈↓ α←␈↓applicable.
␈↓"β␈↓ α←␈↓␈↓ β?Knowledge embedded in the schema-schema assumes that:

␈↓"β␈↓ α←␈↓␈↓ ββ(1)␈↓ β?Data␈α⊃structures␈α⊂have␈α⊃a␈α⊃well-specified␈α⊂syntax.␈α⊃ That␈α⊃is,␈α⊂they
␈↓ α←␈↓␈↓ β?have␈α
a␈α
certain␈α
static␈α
quality␈α
and␈α
maintain␈α
the␈α∞same␈α
structure
␈↓ α←␈↓␈↓ β?and␈α⊂organization␈α⊂over␈α⊂a␈α∂lifetime␈α⊂that␈α⊂includes␈α⊂a␈α⊂number␈α∂of
␈↓ α←␈↓␈↓ β?access,␈α⊂storage,␈α∂and␈α⊂creation␈α∂operations.␈α⊂ One␈α∂obvious␈α⊂set␈α∂of
␈↓ α←␈↓␈↓ β?candidates␈α∞are␈α∂those␈α∞structures␈α∞that␈α∂do␈α∞not␈α∞change␈α∂while␈α∞the
␈↓ α←␈↓␈↓ β?program␈αis␈αexecuting.␈α
 Conversely,␈αapplying␈αthis␈α
to␈αtemporary
␈↓ α←␈↓␈↓ β?structures which are quickly modified would be less successful.

␈↓"β␈↓ α←␈↓␈↓ ββ(2)␈↓ β?Data␈αstructures␈αcan␈αbe␈αspecified␈αin␈αterms␈αof␈αdistinct␈αsub-units,
␈↓ α←␈↓␈↓ β?each␈α⊂of␈α⊃which␈α⊂has␈α⊃a␈α⊂straightforward␈α⊂syntax␈α⊃and␈α⊂is␈α⊃for␈α⊂the
␈↓ α←␈↓␈↓ β?most part independent of the others.
␈↓ α←␈↓␈↓176    KNOWLEDGE ACQUISITION II␈↓ 
∪6-11␈↓

␈↓"β␈↓ α←␈↓␈↓ ββ(3)␈↓ β?Data␈α
structures␈αmay␈α
be␈αinterdependent.␈α
 Thus,␈αpart␈α
of␈αthe␈α
task
␈↓ α←␈↓␈↓ β?of␈α≤specifying␈α≠a␈α≤new␈α≠representation␈α≤is␈α≠to␈α≤describe␈α≠any
␈↓ α←␈↓␈↓ β?interrelationships it may have with other structures.

␈↓"β␈↓ α←␈↓␈↓ ββ(4)␈↓ β?There␈α∩is␈α∩more␈α∪than␈α∩one␈α∩instance␈α∪of␈α∩each␈α∩data␈α∪type.␈α∩ The
␈↓ α←␈↓␈↓ β?utility␈α∀of␈α∪the␈α∀schema␈α∪as␈α∀a␈α∪tool␈α∀for␈α∪dealing␈α∀with␈α∪program
␈↓ α←␈↓␈↓ β?complexity␈αis␈α
dependent␈αon␈α
a␈αuseful␈α
instance-to-schema␈αratio.
␈↓ α←␈↓␈↓ β?If␈αevery␈αdata␈αstructure␈αin␈α
a␈αprogram␈αwere␈αdistinct␈α(a␈α1:1␈α
ratio),
␈↓ α←␈↓␈↓ β?the␈α∞schemata␈α∞would␈α∞offer␈α∞little␈α∞advantage␈α∞in␈α∂knowledge␈α∞base
␈↓ α←␈↓␈↓ β?maintenance.

␈↓ α←␈↓Since␈α∞the␈α∞schemata␈α∞were␈α∞devised␈α∞as␈α
an␈α∞extension␈α∞to␈α∞the␈α∞notion␈α∞of␈α∞a␈α
record
␈↓ α←␈↓structure,␈α∂it␈α∂is␈α∂not␈α∞surprising␈α∂to␈α∂find␈α∂that␈α∞several␈α∂of␈α∂these␈α∂assumptions␈α∞are
␈↓ α←␈↓common to the use of record structures as well.
␈↓ α←␈↓␈↓6-11␈↓ ∧"KNOWLEDGE ABOUT KNOWLEDGE ABOUT REPRESENTATIONS    177␈↓









␈↓"β␈↓ α←␈↓¬␈↓&SCHEMA-SCHEMA␈↓)αβ

␈↓"β␈↓ α←␈↓¬PNTNAME      (ATOM      CREATEIT)
␈↓"β␈↓ α←␈↓¬STRUCT       (PNTNAME   INSLOT)
␈↓"β␈↓ α←␈↓¬PLIST
␈↓"β␈↓ α←␈↓¬       [ (PNTNAME   ((BLANK-INST ADVICE-INST) ASKIT)
␈↓"β␈↓ α←␈↓¬          STRUCT    ((PNTNAME INSLOT)         GIVENIT)
␈↓"β␈↓ α←␈↓¬          PLIST
␈↓"β␈↓ α←␈↓¬                 [ (INSTOF       (( (PNTNAME INSLOT) GIVENIT )            CREATEIT)
␈↓"β␈↓ α←␈↓¬                    DESCR        ((STRING  ASKIT)                         GIVENIT)
␈↓"β␈↓ α←␈↓¬                    AUTHOR       ((ATOM    ASKIT)                         GIVENIT)
␈↓"β␈↓ α←␈↓¬                    DATE         ((INTEGER CREATEIT)                      GIVENIT)
␈↓"β␈↓ α←␈↓¬                    KLEENE       ((SLOTNAME-INST (BLANK-INST ADVICE-INST)) ASKIT))
␈↓"β␈↓ α←␈↓¬                  CREATEIT]
␈↓"β␈↓ α←␈↓¬          FATHER    ( SCHEMA-INST             FINDIT)
␈↓"β␈↓ α←␈↓¬          INSTANCES ( LIST                    ASKIT)
␈↓"β␈↓ α←␈↓¬          STRAN     ( STRING                  FINDIT)
␈↓"β␈↓ α←␈↓¬          INSTOF    ( SCHEMA-SCHEMA           GIVENIT)
␈↓"β␈↓ α←␈↓¬          DESCR     ( STRING                  CREATEIT)
␈↓"β␈↓ α←␈↓¬          AUTHOR    ( ATOM                    ASKIT)
␈↓"β␈↓ α←␈↓¬          DATE      ( INTEGER                 CREATEIT)
␈↓"β␈↓ α←␈↓¬          OFFSPRING ( (KLEENE (0) < SCHEMA-INST >)  ASKIT)
␈↓"β␈↓ α←␈↓¬          RELATIONS ((KLEENE (0) <(UPDATECOM-INST KLEENE (1)
␈↓"β␈↓ α←␈↓¬                                  <(SWITCHCOM-INST KLEENE (1) <KSTRUCT-INST>)>)>)
␈↓"β␈↓ α←␈↓¬                             ASKIT))
␈↓"β␈↓ α←␈↓¬          CREATEIT]
␈↓"β␈↓ α←␈↓¬  FATHER     (SCHEMA-SCHEMA)
␈↓"β␈↓ α←␈↓¬  INSTANCES  ((ALLSCHEMA))
␈↓"β␈↓ α←␈↓¬  STRAN      "knowledge structure"
␈↓"β␈↓ α←␈↓¬  INSTOF     (SCHEMA-SCHEMA)
␈↓"β␈↓ α←␈↓¬  DESCR      "the schema-schema describes the format for all other schemata"
␈↓"β␈↓ α←␈↓¬  AUTHOR     DAVIS
␈↓"β␈↓ α←␈↓¬  DATE       876
␈↓"β␈↓ α←␈↓¬  OFFSPRING  NIL
␈↓"β␈↓ α←␈↓¬  RELATIONS  ((ADDTO (AND* ALLSCHEMA)))




␈↓"β

␈↓"β␈↓ α←␈↓α␈↓ ∧lFig. 6-11.    The schema-schema.    
␈↓"β␈↓ α←␈↓␈↓178    KNOWLEDGE ACQUISITION II␈↓ 
∪6-11␈↓

␈↓"β␈↓ α←␈↓␈↓α6-12    TRACE OF SYSTEM PERFORMANCE:  BUILDING THE
␈↓ α←␈↓α␈↓ β3SCHEMA NETWORK␈↓
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∂example␈α∂demonstrates␈α∂the␈α∂process␈α∞of␈α∂adding␈α∂a␈α∂new␈α∂schema␈α∞to
␈↓ α←␈↓the␈αnetwork␈αin␈αthe␈αvery␈αearly␈αstages␈αof␈αknowledge␈αbase␈αconstruction.␈α At␈αthis
␈↓ α←␈↓point␈α
the␈αsystem␈α
does␈αnot␈α
yet␈α
know␈αanything␈α
about␈αthe␈α
concept␈α
of␈αorganism
␈↓ α←␈↓identity.

␈↓ α←␈↓	***************************************************************

␈↓ α←␈↓	Schema network editor
␈↓ α←␈↓	Command?
␈↓ α←␈↓	++** ␈↓αDOWN␈↓	
␈↓ α←␈↓	    Level 1:
␈↓ α←␈↓	        1) attribute
␈↓ α←␈↓	        2) value of an attribute

␈↓ α←␈↓↓␈↓ βW{There␈αis␈αa␈αsimple␈αeditor␈αthat␈αallows␈αthe␈αuser␈αto␈αexamine␈αthe␈αnetwork
␈↓ α←␈↓↓␈↓ βWand␈α⊂perform␈α⊂various␈α⊂operations␈α⊂on␈α⊂it.␈α⊂ Here␈α⊂he␈α⊂descends␈α⊃from␈α⊂the
␈↓ α←␈↓↓␈↓ βWroot...}

␈↓ α←␈↓	++** ␈↓αSHOW 2␈↓	

␈↓ α←␈↓	VALUE-SCHEMA
␈↓ α←␈↓	------------
␈↓ α←␈↓	  PNTNAME       (ATOM      ASKIT)
␈↓ α←␈↓	  VAL           (PNTNAME   INSLOT)
␈↓ α←␈↓	  PLIST         [(INSTOF  (VALUE-SCHEMA     GIVENIT)
␈↓ α←␈↓	                  DESCR   (STRING           ASKIT)
␈↓ α←␈↓	                  AUTHOR  (ATOM             FINDIT)
␈↓ α←␈↓	                  DATE    (INTEGER          CREATEIT))
␈↓ α←␈↓	                 CREATEIT]
␈↓ α←␈↓	  STRAN         the value of a attribute
␈↓ α←␈↓	  FATHER        (KSTRUCT-SCHEMA)
␈↓ α←␈↓	  OFFSPRING     (SITE-SCHEMA)
␈↓ α←␈↓	  DESCR         the VALUE-SCHEMA describes the format for a
␈↓ α←␈↓	                value of any attribute
␈↓ α←␈↓	  AUTHOR        DAVIS
␈↓ α←␈↓	  DATE          1115
␈↓ α←␈↓	  INSTOF        (SCHEMA-SCHEMA)

␈↓ α←␈↓↓␈↓ βW{examining schemata along the way...}

␈↓ α←␈↓	++** ␈↓αDOWN 2␈↓	
␈↓ α←␈↓	    Level 2:
␈↓ α←␈↓	            2.1) site of a culture

␈↓ α←␈↓	++** ␈↓αADD 2.2␈↓	

␈↓ α←␈↓↓␈↓ βW{until␈αhe␈α
has␈αfound␈α
the␈αright␈αspot␈α
and␈αthen␈α
indicates␈αthat␈α
he␈αwants
␈↓ α←␈↓␈↓6-12␈↓ εaBUILDING THE SCHEMA NETWORK    179␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βWto␈αadd␈αa␈αnew␈α
schema.␈α The␈αsystem␈αretrieves␈αthe␈α
␈↓	SCHEMA-SCHEMA␈↓↓␈αand
␈↓ α←␈↓↓␈↓ βWuses it to guide the process of describing the new data type.}

␈↓ α←␈↓	Please give a phrase to describe this new type of value of an
␈↓ α←␈↓	attribute.
␈↓ α←␈↓	++** ␈↓αIDENTITY OF AN ORGANISM␈↓	

␈↓ α←␈↓	    If there is one word which neatly sums up "identity of an
␈↓ α←␈↓	    organism", please give it (else press RETURN)
␈↓ α←␈↓	    ++** ␈↓αIDENT␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈α
system␈α
needs␈α
a␈α
name␈α
for␈α
the␈α
new␈α
schema␈α
and␈α
asks␈α
the␈α
user␈α
first.
␈↓ α←␈↓↓␈↓ βWIf␈α
he␈α
declines␈αto␈α
supply␈α
one,␈αthe␈α
system␈α
will␈αmanufacture␈α
one␈α
out␈αof
␈↓ α←␈↓↓␈↓ βWthe phrase ``identity of an organism.''}

␈↓ α←␈↓	    Please describe any additional properties of a identity of
␈↓ α←␈↓	    an organism. The answer should be in the form of one or
␈↓ α←␈↓	    more of the following-
␈↓ α←␈↓	        <slot name> <blank specifier> <slot advice>
␈↓ α←␈↓	    [Type 1 set to a line, then an empty line when done.]
␈↓ α←␈↓	    ++** ␈↓αAIR     (KLEENE (1 1 2) <(AIR-INST CF-INST)>)   ASKIT␈↓	
␈↓ α←␈↓	    ++** ␈↓αGRAM       GRAM-INST     ASKIT␈↓	
␈↓ α←␈↓	    ++** ␈↓αMORPH     MORPH-INST     ASKIT␈↓	
␈↓ α←␈↓	    ++** ␈↓αSYNONYM     (KLEENE (1 0) < ATOM >)    ASKIT␈↓	
␈↓ α←␈↓	    ++**

␈↓ α←␈↓↓␈↓ βW{The␈α
user␈α
indicates␈α
several␈αstructural␈α
components␈α
that␈α
are␈α
part␈αof␈α
the
␈↓ α←␈↓↓␈↓ βWnew␈α∪data␈α∀type,␈α∪describing␈α∪them␈α∀in␈α∪the␈α∀standard␈α∪slotname-blank-
␈↓ α←␈↓↓␈↓ βWadvice␈αformat.␈α These␈α
are␈αin␈αaddition␈α
to␈αthe␈αstructural␈α
conventions␈αit
␈↓ α←␈↓↓␈↓ βWinherits by virtue of being a type of ␈↓	VALUE␈↓↓.}

␈↓ α←␈↓	    Sorry, but the following are invalid -
␈↓ α←␈↓	        SYNONYM is not a known <slot name>
␈↓ α←␈↓	    Please answer again [use the same answer if you really
␈↓ α←␈↓	    meant it.]

␈↓ α←␈↓	    ++** ␈↓αSYNONYM (KLEENE (1 0) < ATOM >) ASKIT␈↓	

␈↓ α←␈↓↓␈↓ βW{The␈αconcept␈αof␈αa␈αsynonym␈αis␈αas␈αyet␈αunknown␈αto␈αthe␈αsystem;␈αso␈αit␈αtoo
␈↓ α←␈↓↓␈↓ βWhas␈α∃to␈α∃be␈α∃described.␈α∃ This␈α∃is␈α∃set␈α∃up␈α∃as␈α∃a␈α∃subproblem,␈α∃and␈α∀the
␈↓ α←␈↓↓␈↓ βW␈↓	SLOTNAME-SCHEMA␈↓↓ is used to guide the description.}

␈↓ α←␈↓	      Please tell me a few things about the concept of SYNONYM
␈↓ α←␈↓	      as a <slot name>.

␈↓ α←␈↓	        Please give a short phrase which can be used to ask for
␈↓ α←␈↓	        the contents of this slot.
␈↓ α←␈↓	        [type an empty line when done]
␈↓ α←␈↓	        ++** ␈↓αPLEASE GIVE ALL SYNONYMS OR ABBREVIATIONS␈↓	
␈↓ α←␈↓	        ++** ␈↓αFOR * WHICH YOU WOULD LIKE THE SYSTEM TO␈↓	
␈↓"β␈↓ α←␈↓␈↓180    KNOWLEDGE ACQUISITION II␈↓ 
∪6-12␈↓

␈↓"β␈↓ α←␈↓	        ++** ␈↓αACCEPT.␈↓	
␈↓ α←␈↓	        ++**

␈↓ α←␈↓↓␈↓ βW{Recall␈αthat␈αthe␈αasterisk␈αis␈αused␈α
to␈αindicate␈αa␈αgap␈αto␈αbe␈αfilled␈α
in␈αthe
␈↓ α←␈↓↓␈↓ βWtemplate.␈α
 In␈α
this␈α
case␈α∞it␈α
will␈α
be␈α
filled␈α
in␈α∞with␈α
the␈α
name␈α
of␈α∞the␈α
new
␈↓ α←␈↓↓␈↓ βWidentity being acquired.}

␈↓ α←␈↓	        Please give a short phrase which can be used to display
␈↓ α←␈↓	        the contents of this slot.
␈↓ α←␈↓	        [type an empty line when done]
␈↓ α←␈↓	        ++** ␈↓αTHE SYNONYMS OF * ARE␈↓	
␈↓ α←␈↓	        ++**

␈↓ α←␈↓	        Please give a description of SYNONYM.
␈↓ α←␈↓	        [type an empty line when done]
␈↓ α←␈↓	        ++** ␈↓αSINCE MANY ORGANISM NAMES ARE LONG AND␈↓	
␈↓ α←␈↓	        ++** ␈↓αUNWIELDY, SHORTER SYNONYMS ARE OFTEN␈↓	
␈↓ α←␈↓	        ++** ␈↓αUSED. THOSE SYNONYMS ARE PART OF THE␈↓	
␈↓ α←␈↓	        ++** ␈↓αDATA STRUCTURE WHICH REPRESENTS AN␈↓	
␈↓ α←␈↓	        ++** ␈↓αORGANISM IDENTITY.␈↓	
␈↓ α←␈↓	        ++**

␈↓ α←␈↓	        Please edit and complete this skeleton function
␈↓ α←␈↓	        definition for the SYNONYM-EXPERT:

␈↓ α←␈↓	[NLAMBDA (BLANKS ADVICE)
␈↓ α←␈↓	  (SELECTQ ADVICE
␈↓ α←␈↓	           (ASKIT (ASK-XPERT BLANKS (QUOTE SYNONYM)))
␈↓ α←␈↓	           (GIVENIT BLANKS)
␈↓ α←␈↓	           (FINDIT  )
␈↓ α←␈↓	           (CREATEIT )
␈↓ α←␈↓	           (INSLOT (APPLY* (GETEXPERT BLANKS) SCHEMA
␈↓ α←␈↓	                                              (QUOTE GETALL)))
␈↓ α←␈↓	           [GETONE (CAR (GETP KSNAME (QUOTE SYNONYM]
␈↓ α←␈↓	           (GETALL (GETP KSNAME (QUOTE SYNONYM)))
␈↓ α←␈↓	           (GETNEXT (NEXTONE SCHEMA (QUOTE SYNONYM))
␈↓ α←␈↓	           (NOADVICE BLANKS (QUOTE SYNONYM-EXPERT]
␈↓ α←␈↓	        tty:
␈↓ α←␈↓	        *

␈↓ α←␈↓↓␈↓ βW{Since␈α∀there␈α∀is␈α∀a␈α∀slot-expert␈α∀associated␈α∀with␈α∀every␈α∀slotname,␈α∀the
␈↓ α←␈↓↓␈↓ βWacquisition␈α⊂of␈α⊂the␈α⊂slot-expert␈α⊂becomes␈α⊂a␈α⊂new␈α⊂sub-task.␈α⊃ Recall␈α⊂that
␈↓ α←␈↓↓␈↓ βWeven␈αthough␈αit␈αis␈αa␈αfunction,␈αthe␈αslot-expert␈αis␈αviewed␈αfor␈αthe␈αmoment
␈↓ α←␈↓↓␈↓ βWas␈α
simply␈α∞another␈α
data␈α∞type,␈α
one␈α∞of␈α
whose␈α∞components␈α
is␈α∞a␈α
function
␈↓ α←␈↓↓␈↓ βWdefinition.␈α
 All␈α
of␈α
the␈α
other␈α
components␈α
are␈α
sufficiently␈α
stylized␈α
that
␈↓ α←␈↓↓␈↓ βWthey␈αcan␈α
be␈αmanufactured␈α
by␈αthe␈α
system␈αitself,␈α
and␈αthis␈αoccurs␈α
without
␈↓ α←␈↓↓␈↓ βWaid from the user.
␈↓ α←␈↓↓␈↓ βWThe␈αfunction␈αdefinition␈αis␈αcomplex␈αenough␈αto␈αbe␈αan␈αexception␈αto␈αthis,
␈↓ α←␈↓↓␈↓ βWbut␈αeven␈αhere␈αthere␈αis␈αenough␈αstylization␈αthat␈αthe␈αsystem␈αcan␈αprepare
␈↓ α←␈↓␈↓6-12␈↓ εaBUILDING THE SCHEMA NETWORK    181␈↓

␈↓"β␈↓ α←␈↓↓␈↓ βWa␈αuseful␈αskeleton␈αto␈αbe␈αcompleted␈αby␈αthe␈αuser.␈α The␈αstandard␈α␈↓¬INTERLISP␈↓↓
␈↓ α←␈↓↓␈↓ βWeditor␈α∩is␈α∩invoked␈α∩(announcing␈α∩itself␈α∩with␈α∩the␈α∩``␈↓	tty:␈↓↓''␈α∩prompt),␈α⊃to
␈↓ α←␈↓↓␈↓ βWallow␈α
the␈αuser␈α
to␈αmake␈α
any␈αnecessary␈α
changes.␈α Since␈α
not␈α
every␈αpiece
␈↓ α←␈↓↓␈↓ βWof␈αadvice␈αmakes␈αsense␈αfor␈α
every␈αslot-expert,␈αthe␈αuser␈αmay␈α
delete␈αsome
␈↓ α←␈↓↓␈↓ βWof␈α↔the␈α⊗entries.␈α↔ Other␈α↔entries␈α⊗may␈α↔be␈α⊗expanded␈α↔to␈α↔account␈α⊗for
␈↓ α←␈↓↓␈↓ βWadditional␈α∂representation␈α∂conventions,␈α∂or␈α∂edited␈α∂because␈α∂the␈α∞original
␈↓ α←␈↓↓␈↓ βWskeleton␈α∂is␈α∂at␈α∞best␈α∂a␈α∂rough␈α∞guess.␈α∂ The␈α∂point␈α∞in␈α∂having␈α∂the␈α∞system
␈↓ α←␈↓↓␈↓ βWproduce␈α
the␈αskeleton␈α
is␈α
not␈αto␈α
automate␈α
the␈αcreation␈α
of␈α
code,␈αbut␈α
rather
␈↓ α←␈↓↓␈↓ βWto␈α∃make␈α∃it␈α∃as␈α∃easy␈α∃as␈α∃possible␈α∃for␈α∃the␈α∃user␈α∃to␈α∃supply␈α∃all␈α∃the
␈↓ α←␈↓↓␈↓ βWinformation that the system will eventually need.}

␈↓ α←␈↓	        .
␈↓ α←␈↓	        .
␈↓ α←␈↓	        .

␈↓ α←␈↓	        * ␈↓αOK␈↓	

␈↓ α←␈↓↓␈↓ βW{The user finishes the editing (which has been omitted here).}

␈↓ α←␈↓	        Done with the concept of SYNONYM as a <slot name> now.

␈↓ α←␈↓↓␈↓ βW{Having␈αfinally␈α
finished␈αwith␈α
the␈αnew␈α
slotname,␈αthe␈α
dialog␈αreturns␈α
to
␈↓ α←␈↓↓␈↓ βWthe last item needed for the new schema ...}

␈↓ α←␈↓	    Please specify all updating to other data structures which
␈↓ α←␈↓	    will be necessary when a new instance of a identity of an
␈↓ α←␈↓	    organism is acquired.  The answer should be in the form of
␈↓ α←␈↓	    one or more of the following-
␈↓ α←␈↓	    <update command> [1 or more: <selection command>
␈↓ α←␈↓	                               [1 or more: <data structure>]]
␈↓ α←␈↓	    [Type 1 set to a line, then an empty line when done.]
␈↓ α←␈↓	    ++** ␈↓αADDTO (AND* ORGANISMS)␈↓	
␈↓ α←␈↓	    ++**

␈↓ α←␈↓	    Ok, finished defining IDENT-SCHEMA.

␈↓ α←␈↓	  Level 2:
␈↓ α←␈↓	            2.1) site of a culture
␈↓ α←␈↓	            2.2) identity of an organism
␈↓ α←␈↓	Command?
␈↓ α←␈↓	++**

␈↓ α←␈↓↓␈↓ βW{... and then is done.}

␈↓"β␈↓ α←␈↓␈↓α6-12-1    Comments on the trace␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αsystem␈αrequires␈αonly␈αa␈αvery␈αsmall␈αcore␈αof␈αknowledge␈αas␈αthe␈αbasis
␈↓ α←␈↓for␈αthe␈αschema␈α
network␈αconstruction␈αshown␈αin␈α
this␈αtrace.␈α In␈αaddition␈α
to␈αthe
␈↓ α←␈↓network␈α
editor␈αand␈α
the␈α
schema␈αinterpreter,␈α
it␈α
requires␈αonly␈α
five␈αschemata␈α
and
␈↓ α←␈↓␈↓182    KNOWLEDGE ACQUISITION II␈↓ 
∪6-12␈↓

␈↓"β␈↓ α←␈↓a␈α
small␈α
number␈α
of␈α
instances.␈↓
16␈↓␈α
From␈α
this␈α
core␈α
of␈α
knowledge␈α∞everything␈α
else
␈↓ α←␈↓can␈αbe␈αbuilt.␈α As␈αa␈αdemonstration,␈αthe␈αnetwork␈αshown␈αabove␈αin␈αFig.␈α6-3␈αwas
␈↓ α←␈↓constructed␈αin␈αthis␈αfashion.␈α The␈αsingle␈αprocess␈αof␈αschema␈αinterpretation␈αwas
␈↓ α←␈↓used␈α→to␈α→guide␈α_the␈α→construction␈α→of␈α_the␈α→base␈α→of␈α_representation-specific
␈↓ α←␈↓knowledge␈α
and␈α∞then␈α
used␈α∞to␈α
instantiate␈α∞it␈α
in␈α∞order␈α
to␈α∞build␈α
a␈α∞small␈α
object-
␈↓ α←␈↓level␈α
knowledge␈α
base.␈α
 The␈α
system␈αwas␈α
thus␈α
bootstrapped␈α
from␈α
the␈αschema-
␈↓ α←␈↓schema and a few associated structures.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α
practice,␈α∞the␈α
content␈α
of␈α∞the␈α
knowledge␈α
base␈α∞would␈α
not␈α∞already␈α
be
␈↓ α←␈↓determined,␈α⊗so␈α∃the␈α⊗process␈α∃would␈α⊗proceed␈α∃slightly␈α⊗differently.␈α⊗ A␈α∃basic
␈↓ α←␈↓skeleton␈α
of␈αthe␈α
schema␈α
network␈αshould␈α
be␈α
constructed␈αfirst,␈α
using␈αthe␈α
network
␈↓ α←␈↓editor.␈α
 After␈αa␈α
few␈α
major␈αbranches␈α
have␈α
been␈αsupplied,␈α
it␈α
is␈αthen␈α
convenient
␈↓ α←␈↓to␈α⊃go␈α⊃back␈α⊃to␈α⊃typing␈α⊃in␈α⊃new␈α⊃rules␈α⊃and␈α⊃to␈α⊃allow␈α⊃the␈α⊃system␈α⊃to␈α⊃guide␈α⊂the
␈↓ α←␈↓necessary␈α
network␈α
growth.␈α
 As␈α
in␈α
the␈α
example␈α
of␈α
acquiring␈α
a␈α
new␈αattribute,
␈↓ α←␈↓this␈αmeans␈α
that␈αa␈αnew␈α
rule␈αmight␈α
trigger␈αthe␈αaddition␈α
of␈αa␈α
new␈αbranch␈αto␈α
the
␈↓ α←␈↓network␈α(the␈αnew␈αdata␈αtype)␈αand␈αthen␈αtrigger␈αseveral␈αinstantiations␈αof␈αit␈α(the
␈↓ α←␈↓nutrients).
␈↓"β␈↓ α←␈↓␈↓ β?While␈α⊂rules␈α∂can␈α⊂be␈α⊂entered␈α∂from␈α⊂the␈α∂very␈α⊂beginning␈α⊂of␈α∂knowledge
␈↓ α←␈↓base␈αconstruction,␈αthis␈αtends␈αinitially␈αto␈αproduce␈αdialogs␈αthat␈αare␈αdifficult␈αfor
␈↓ α←␈↓the␈αuser␈αto␈αfollow.␈α With␈αan␈αempty␈αschema␈αnetwork,␈αthe␈αfirst␈αline␈αof␈αthe␈αfirst
␈↓ α←␈↓rule␈α⊃will␈α⊃trigger␈α⊂a␈α⊃long␈α⊃and␈α⊂deeply␈α⊃recursive␈α⊃dialog.␈α⊂ In␈α⊃general,␈α⊃early␈α⊂in
␈↓ α←␈↓knowledge␈α∩base␈α∩construction,␈α∩the␈α∩smallest␈α∩addition␈α∩tends␈α∩to␈α∪trigger␈α∩many
␈↓ α←␈↓other␈αadditions.␈α It␈αis␈αeasiest␈αto␈αstart␈αby␈αbuilding␈αa␈αbasic␈αnetwork␈αa␈αpiece␈αat␈αa
␈↓ α←␈↓time with the editor.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈α∞the␈α∞network␈α
can␈α∞conceivably␈α∞grow␈α
quite␈α∞complex,␈α∞some␈α
simple
␈↓ α←␈↓heuristics␈αhave␈α
been␈αembedded␈α
in␈αthe␈α
editor.␈α To␈α
help␈αdeal␈α
with␈αthe␈α
problem
␈↓ α←␈↓of␈α∃potential␈α∃interconnections␈α⊗of␈α∃a␈α∃new␈α⊗schema,␈α∃the␈α∃editor␈α⊗can␈α∃propose
␈↓ α←␈↓candidates.␈α If␈αa␈αnew␈αschema␈αwere␈αadded␈αin␈αthe␈αthird␈αlevel␈αof␈αthe␈αnetwork␈α
of
␈↓ α←␈↓Fig. 6-3, the editor would suggest:

␈↓"β␈↓ α←␈↓	Listed below are 1 or more possible sub-classifications of
␈↓"β␈↓ α←␈↓	this new concept. Please indicate [Y or N] each one that
␈↓"β␈↓ α←␈↓	applies.
␈↓"β␈↓ α←␈↓	  1 - an attribute whose value is "true" or "false"
␈↓"β␈↓ α←␈↓	  ++** ␈↓αYES␈↓	
␈↓"β␈↓ α←␈↓	  2 - a multi-valued attribute
␈↓"β␈↓ α←␈↓	  ++** ␈↓αYES␈↓	
␈↓"β␈↓ α←␈↓	  3 - a single-valued attribute
␈↓"β␈↓ α←␈↓	  ++** ␈↓αYES␈↓	

␈↓"β␈↓ α←␈↓α␈↓ ∧↔Fig. 6-12.    Proposing potential connections.    

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[16]␈α∂Since␈α⊂the␈α∂schema-schema␈α∂needs␈α⊂to␈α∂refer␈α∂to␈α⊂the␈α∂concepts␈α⊂of␈α∂slotnames,
␈↓ α←␈↓slot-experts,␈αadvice,␈α
and␈αblanks,␈αthe␈α
schemata␈αfor␈αthese␈α
must␈αbe␈αsupplied␈α
and
␈↓ α←␈↓cannot␈α∂be␈α∂bootstrapped.␈α∂ The␈α∞instantiations␈α∂required␈α∂are␈α∂the␈α∂slotname␈α∞and
␈↓ α←␈↓slot-experts␈αfor␈α
each␈αof␈αthe␈α
slotnames␈αfound␈αin␈α
a␈αschema,␈α
and␈αinstantiations
␈↓ α←␈↓of the advice schema for the nine pieces of advice.
␈↓ α←␈↓␈↓6-12␈↓ εaBUILDING THE SCHEMA NETWORK    183␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂editor␈α⊂examines␈α⊃the␈α⊂siblings␈α⊂of␈α⊂the␈α⊃new␈α⊂schema␈α⊂and␈α⊃notes␈α⊂to
␈↓ α←␈↓which␈α∞other␈α∞schemata␈α∞they␈α∞are␈α∞connected.␈α∞ If␈α∞a␈α∞sufficient␈α∞percentage␈α∞of␈α
the
␈↓ α←␈↓siblings␈α⊗share␈α⊗a␈α∃common␈α⊗offspring,␈α⊗that␈α∃offspring␈α⊗becomes␈α⊗a␈α∃potential
␈↓ α←␈↓connection␈αin␈α
the␈αschema␈αnetwork.␈α
 In␈αthis␈αcase␈α
the␈αeditor␈α(correctly)␈α
proposes
␈↓ α←␈↓the three schemata to which all the other five siblings are attached.
␈↓"β␈↓ α←␈↓␈↓ β?An␈αanalogous␈α
sort␈αof␈αaid␈α
is␈αavailable␈αwhen␈α
specifying␈αthe␈αstructure␈α
of
␈↓ α←␈↓a␈α
new␈α
data␈αtype.␈α
 The␈α
network␈αeditor␈α
examines␈α
the␈α
three␈αstructure-defining
␈↓ α←␈↓slots␈α∞(the␈α
print␈α∞name,␈α
value,␈α∞and␈α
property␈α∞list)␈α
of␈α∞the␈α
new␈α∞schema's␈α
siblings
␈↓ α←␈↓and␈α∩detects␈α∩regularities␈α∩in␈α⊃a␈α∩manner␈α∩similar␈α∩to␈α⊃the␈α∩way␈α∩rule␈α∩models␈α⊃are
␈↓ α←␈↓created.␈α These␈αare␈αthen␈α
displayed␈αto␈αthe␈αuser␈α
and,␈αlike␈αthe␈αrule␈α
models,␈αcan
␈↓ α←␈↓be useful reminders of overlooked details.

␈↓"β␈↓ α←␈↓␈↓α6-13    LEVELS OF KNOWLEDGE␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃mechanisms␈α⊃reviewed␈α∩above␈α⊃provide␈α⊃an␈α⊃extensive␈α∩amount␈α⊃of
␈↓ α←␈↓machinery␈α∀for␈α∀encoding␈α∪knowledge␈α∀about␈α∀representations.␈α∪ But␈α∀it␈α∀is␈α∪not
␈↓ α←␈↓enough␈α∂simply␈α⊂to␈α∂provide␈α∂the␈α⊂machinery--if␈α∂the␈α∂result␈α⊂is␈α∂to␈α⊂be␈α∂something
␈↓ α←␈↓more␈α
than␈α∞``yet␈α
another␈α
knowledge␈α∞representation␈α
formalism,''␈α
there␈α∞must␈α
be
␈↓ α←␈↓some␈αsense␈αof␈αorganization␈αand␈αmethodology␈αthat␈αsuggests␈αhow␈αall␈αthis␈αought
␈↓ α←␈↓to be used.
␈↓"β␈↓ α←␈↓␈↓ β?Organization␈αis␈αprovided␈αby␈αa␈αcommon␈αtheme␈αthat␈αserves␈αto␈αunify␈αall
␈↓ α←␈↓of␈α∂the␈α∞proposed␈α∂mechanisms:␈α∞the␈α∂notion␈α∞of␈α∂␈↓↓levels␈α∞of␈α∂knowledge␈↓.␈α∂ There␈α∞are
␈↓ α←␈↓several␈α⊂different␈α∂(and␈α⊂independent)␈α∂stratifications␈α⊂of␈α∂knowledge␈α⊂implicit␈α∂in
␈↓ α←␈↓the formalism developed above.  Two of the most important involve:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?describing␈α
knowledge␈α
in␈α
the␈αsystem␈α
at␈α
different␈α
levels␈αof␈α
detail,
␈↓ α←␈↓␈↓ β?and
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?classifying it according to its level of generality.

␈↓ α←␈↓In␈α
both␈αcases,␈α
the␈α
important␈αcontribution␈α
is␈α
a␈αframework␈α
for␈α
organizing␈αthe
␈↓ α←␈↓relevant␈α⊂knowledge␈α⊂about␈α⊂representations.␈α⊂ The␈α⊂idea␈α⊂of␈α⊂different␈α⊃levels␈α⊂of
␈↓ α←␈↓detail␈αindicates␈αthat␈αrepresentations␈α(e.g.,␈α␈↓↓value␈↓␈αor␈α␈↓↓attribute␈↓)␈αcan␈α
be␈αdescribed
␈↓ α←␈↓at␈α∂the␈α∂level␈α∂of␈α∂global␈α∂organization␈α∂(as␈α∂in␈α∂the␈α∂schema␈α∂hierarchy),␈α∂of␈α∂logical
␈↓ α←␈↓structure␈α∩(as␈α⊃in␈α∩the␈α⊃schemata),␈α∩and␈α⊃of␈α∩implementation␈α⊃(as␈α∩in␈α⊃information
␈↓ α←␈↓associated␈α⊗with␈α∃the␈α⊗slotnames).␈α∃ These␈α⊗levels␈α∃provide␈α⊗an␈α∃organizational
␈↓ α←␈↓scheme␈αthat␈αmakes␈αit␈αeasier␈αto␈αspecify␈αand␈αto␈αkeep␈αtrack␈αof␈αthe␈αlarge␈αstore␈αof
␈↓ α←␈↓information␈α∩about␈α∪representations␈α∩required␈α∩by␈α∪the␈α∩acquisition␈α∪task.␈α∩ The
␈↓ α←␈↓different␈α∃levels␈α∀of␈α∃generality␈α∀for␈α∃classifying␈α∀knowledge␈α∃include: ␈α∀domain
␈↓ α←␈↓specific,␈α∨representation␈α∨specific,␈α∨and␈α∨representation␈α∨independent.␈α≡ As
␈↓ α←␈↓explained␈α⊃below,␈α⊃the␈α⊃idea␈α⊃of␈α⊃maintaining␈α⊃clear␈α⊃distinctions␈α∩between␈α⊃these
␈↓ α←␈↓different␈α⊗kinds␈α⊗of␈α⊗knowledge␈α⊗is␈α⊗an␈α⊗important␈α⊗contributor␈α⊗to␈α↔much␈α⊗of
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓'s current range of application.

␈↓"β␈↓ α←␈↓␈↓α6-13-1    Level of detail␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈α
noted␈α
in␈α
Section␈α
6-8,␈α
the␈α
schema␈α
hierarchy,␈α∞individual␈α
schemata,
␈↓ α←␈↓and␈α∨slotnames␈α∨each␈α∨encode␈α their␈α∨own␈α∨form␈α∨of␈α knowledge␈α∨about
␈↓ α←␈↓␈↓184    KNOWLEDGE ACQUISITION II␈↓ 
∪6-13␈↓

␈↓"β␈↓ α←␈↓representations.␈α≥ The␈α≥hierarchy␈α≡indicates␈α≥the␈α≥global␈α≡organization␈α≥of
␈↓ α←␈↓representations␈α∃in␈α∃the␈α⊗system␈α∃and␈α∃provides␈α⊗a␈α∃foundation␈α∃for␈α⊗both␈α∃the
␈↓ α←␈↓acquisition␈α∪of␈α∪new␈α∪instances␈α∪of␈α∩existing␈α∪primitives␈α∪(a␈α∪process␈α∪of␈α∩descent
␈↓ α←␈↓through␈α∂the␈α∂hierarchy␈α∂and␈α∂instantiation␈α∂of␈α∂the␈α∂schemata␈α∂encountered)␈α∂and
␈↓ α←␈↓the␈αacquisition␈αof␈α
new␈αkinds␈αof␈α
primitives␈α(a␈αprocess␈α
of␈αadding␈αnew␈α
branches
␈↓ α←␈↓to␈α∞the␈α∞hierarchy).␈↓
17␈↓␈α∞The␈α∞schemata␈α∞describe␈α∞the␈α∞logical␈α∞structure␈α∞and␈α∞logical
␈↓ α←␈↓interrelationships␈αof␈αindividual␈αrepresentations␈αand,␈αas␈αprototypes,␈αprovide␈αa
␈↓ α←␈↓focus␈α⊗for␈α↔the␈α⊗organization␈α↔of␈α⊗knowledge␈α↔about␈α⊗a␈α↔representation.␈α⊗ The
␈↓ α←␈↓slotnames␈α#have␈α"associated␈α#with␈α"them␈α#information␈α#concerning␈α"the
␈↓ α←␈↓implementation␈α∀of␈α∃a␈α∀specific␈α∃representation,␈α∀information␈α∃at␈α∀the␈α∃level␈α∀of
␈↓ α←␈↓programming-language␈α→constructs␈α→and␈α_conventions␈α→(e.g.,␈α→variable␈α_name
␈↓ α←␈↓uniqueness, etc.).

␈↓"β␈↓ α←␈↓␈↓α6-13-2    Level of generality␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Much␈α∀of␈α∪the␈α∀range␈α∀of␈α∪applicability␈α∀of␈α∪␈↓¬TEIRESIAS␈↓␈α∀results␈α∀from␈α∪the
␈↓ α←␈↓isolation␈α⊃and␈α⊂stratification␈α⊃of␈α⊂the␈α⊃three␈α⊂kinds␈α⊃of␈α⊂knowledge␈α⊃shown␈α⊂below.
␈↓ α←␈↓The␈αbase␈αof␈α␈↓↓domain-specific␈↓␈αknowledge␈αat␈αlevel␈α0␈αconsists␈αof␈αthe␈αcollection␈αof
␈↓ α←␈↓all instances of each representation.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂base␈α∂of␈α∂␈↓↓representation-specific␈↓␈α∂knowledge␈α∂at␈α∂level␈α∂1␈α∂consists␈α∞of
␈↓ α←␈↓the␈α⊃schemata,␈α⊃which␈α⊃are,␈α⊃in␈α∩effect,␈α⊃the␈α⊃declarations␈α⊃of␈α⊃the␈α∩extended␈α⊃data
␈↓ α←␈↓types.␈α⊃ These␈α⊃have␈α⊃a␈α⊃degree␈α⊃of␈α⊃domain␈α⊃independence␈α⊃since␈α∩they␈α⊃describe
␈↓ α←␈↓what␈α∃an␈α∃attribute␈α∃is,␈α∃what␈α∃a␈α∃value␈α∃is,␈α∃etc.,␈α∃without␈α∃requiring␈α∃␈↓↓a␈α∀priori␈↓
␈↓ α←␈↓knowledge of the domain in which those descriptions will be instantiated.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩base␈α∩of␈α∩␈↓↓representation-independent␈↓␈α∩knowledge␈α∩at␈α∩level␈α⊃2--the
␈↓ α←␈↓schema-schema--describes␈α
what␈αa␈α
declaration␈α
looks␈αlike.␈α
 At␈α
this␈αlevel␈α
resides
␈↓ α←␈↓knowledge␈α↔about␈α↔representations␈α↔in␈α_general␈α↔and␈α↔about␈α↔the␈α_process␈α↔of
␈↓ α←␈↓specifying them via declarations.

␈↓"β␈↓ α←␈↓␈↓ ββ(0)␈↓ β?The knowledge base of the performance program contains:
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?␈↓↓object-level␈↓ knowledge that is
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?␈↓↓domain specific␈↓ and is formed by
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?instantiating␈α
the␈αappropriate␈α
schema␈αto␈α
form␈αa␈α
␈↓↓new␈αinstance␈α
of
␈↓ α←␈↓↓␈↓ βoan existing conceptual primitive␈↓.

␈↓"β␈↓ α←␈↓␈↓ ββ(1)␈↓ β?The knowledge about representations (the schemata) contains:
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?␈↓↓meta-level␈↓ knowledge that is
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?␈↓↓representation specific␈↓ and is formed by
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?instantiating␈α_the␈α↔schema-schema␈α_to␈α↔form␈α_a␈α↔␈↓↓new␈α_type␈α↔of
␈↓ α←␈↓↓␈↓ βoconceptual primitive␈↓.

␈↓"β␈↓ α←␈↓␈↓ ββ(2)␈↓ β?The schema-schema contains:
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[17]␈α∂Note␈α∞that␈α∂the␈α∞entire␈α∂schema␈α∞hierarchy␈α∂is␈α∞viewed␈α∂here␈α∞as␈α∂dealing␈α∞with
␈↓ α←␈↓information␈α~at␈α~a␈α~single␈α≠level␈α~of␈α~detail␈α~(viz.,␈α~global␈α≠organization␈α~of
␈↓ α←␈↓representations).␈α∂Viewed␈α∂by␈α∂itself,␈α∂it␈α∞is␈α∂of␈α∂course␈α∂yet␈α∂another␈α∞(independent)
␈↓ α←␈↓structuring of knowledge in the system into various levels.
␈↓ α←␈↓␈↓6-13␈↓ πZLEVELS OF KNOWLEDGE    185␈↓

␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?␈↓↓second order meta-level␈↓ knowledge that is
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?␈↓↓representation independent␈↓ and is formed by
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?hand.

␈↓"β␈↓ α←␈↓α␈↓ αxFig. 6-13.    Levels of generality of knowledge about representations.    

␈↓"β␈↓ α←␈↓␈↓ β?While␈αlevel␈α2␈αis␈αformed␈αby␈αhand,␈αit␈αis␈αthe␈αonly␈αbody␈αof␈αknowledge␈αin
␈↓ α←␈↓the␈α
system␈α
for␈α
which␈α
this␈α
is␈α
true,␈α
and␈α
it␈α
forms␈α
a␈α
small␈α
core␈α
of␈α
knowledge␈α
from
␈↓ α←␈↓which␈α∩everything␈α∪else␈α∩can␈α∪be␈α∩built.␈α∩ For␈α∪example,␈α∩the␈α∪schema␈α∩hierarchy
␈↓ α←␈↓shown␈α_in␈α_Fig.␈α_6-3␈α_(and␈α↔all␈α_associated␈α_structures)␈α_was␈α_constructed␈α↔by
␈↓ α←␈↓bootstrapping␈α∞from␈α∞the␈α∞schema-schema␈α∞and␈α∞the␈α∞core␈α∞of␈α∞structures␈α∂noted␈α∞in
␈↓ α←␈↓Section␈α⊃6-12-1.␈α⊃ The␈α⊃single␈α⊃process␈α⊃of␈α⊃schema␈α⊃interpretation␈α⊃was␈α⊃used␈α⊃to
␈↓ α←␈↓guide␈α
the␈α∞construction␈α
of␈α
the␈α∞base␈α
of␈α
representation-specific␈α∞knowledge␈α
(the
␈↓ α←␈↓hierarchy␈α⊃and␈α⊂schemata)␈α⊃and␈α⊂then␈α⊃used␈α⊂to␈α⊃instantiate␈α⊂it␈α⊃to␈α⊂build␈α⊃a␈α⊂small
␈↓ α←␈↓object-level knowledge base.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α
reason␈α
that␈α
this␈α
is␈α
a␈αpractical␈α
approach␈α
is␈α
the␈α
great␈α
leverage␈αin
␈↓ α←␈↓the␈α∞notion␈α∞of␈α
a␈α∞schema␈α∞as␈α∞a␈α
prototype.␈α∞ The␈α∞current␈α∞performance␈α
program,
␈↓ α←␈↓for␈α⊃instance,␈α∩contains␈α⊃knowledge␈α∩about␈α⊃some␈α⊃125␈α∩organisms,␈α⊃but␈α∩a␈α⊃single
␈↓ α←␈↓schema␈α∃serves␈α∃to␈α∃characterize␈α∃every␈α∀one␈α∃of␈α∃them.␈α∃ There␈α∃are␈α∃some␈α∀25
␈↓ α←␈↓different␈αrepresentations␈αin␈α
the␈αprogram,␈αrequiring␈α
25␈αschemata;␈αyet␈α
a␈αsingle
␈↓ α←␈↓schema-schema serves to characterize all of them.
␈↓"β␈↓ α←␈↓␈↓ β?It␈αwas,␈αin␈αfact,␈αprecisely␈αsuch␈αutilitarian␈αconsiderations␈αthat␈αmotivated
␈↓ α←␈↓the␈α⊂initial␈α⊂creation␈α⊃of␈α⊂the␈α⊂schema-schema.␈α⊃ Recall␈α⊂that␈α⊂the␈α⊃schemata␈α⊂were
␈↓ α←␈↓developed␈αbecause␈αthere␈αwere␈αmany␈αdetails␈αinvolved␈αin␈αcreating␈αa␈αnew␈α
object
␈↓ α←␈↓and␈α
adding␈α
it␈α
to␈α
the␈αsystem.␈α
 But␈α
there␈α
turned␈α
out␈αto␈α
be␈α
a␈α
large␈α
number␈αof
␈↓ α←␈↓details␈α⊂involved␈α⊂in␈α∂creating␈α⊂all␈α⊂the␈α∂necessary␈α⊂schemata,␈α⊂too.␈α⊂ The␈α∂schema-
␈↓ α←␈↓schema␈α
was␈α
thus␈α
the␈α
result␈α
of␈α
the␈α
straightforward␈α
recursive␈α
application␈αof␈α
the
␈↓ α←␈↓basic idea, for precisely the same reason.

␈↓"β␈↓ α←␈↓␈↓α6-13-3    Impact␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αdirect␈αadvantages␈αof␈αthis␈αstratification␈αarise␈αfrom␈αthe␈αcapabilities
␈↓ α←␈↓it␈αsupports.␈α
 The␈αcompartmentalization␈αof␈α
knowledge␈αsuggested␈αby␈α
the␈αlevels
␈↓ α←␈↓of␈αgenerality,␈αfor␈αinstance,␈αprovides␈α
an␈αincreased␈αrange␈αof␈αapplicability␈αof␈α
the
␈↓ α←␈↓system.␈α The␈α
single␈αschema-instantiation␈αprocess␈α
can␈αbe␈αused␈α
with␈αthe␈αcore␈α
of
␈↓ α←␈↓representation-specific␈α
knowledge␈α
in␈α
a␈α
range␈α
of␈α
different␈α
domains,␈α
or␈α
it␈αcan
␈↓ α←␈↓be␈α∩used␈α∩with␈α∩the␈α∩representation-independent␈α∩knowledge␈α∩over␈α∩a␈α∪range␈α∩of
␈↓ α←␈↓representations.␈α⊂ Describing␈α⊂representations␈α⊂at␈α⊂different␈α⊂levels␈α⊂of␈α⊂detail,␈α∂on
␈↓ α←␈↓the␈αother␈αhand,␈α
offers␈αa␈αframework␈αfor␈α
organizing␈αand␈αkeeping␈αtrack␈α
of␈αthe
␈↓ α←␈↓required␈α∞information.␈α∂ It␈α∞also␈α∞provides␈α∂a␈α∞useful␈α∞degree␈α∂of␈α∞flexibility␈α∂in␈α∞the
␈↓ α←␈↓system,␈α
because␈α
the␈α
multiple␈α
levels␈αof␈α
description␈α
insulate␈α
changes␈α
at␈αone␈α
level
␈↓ α←␈↓from␈α↔the␈α↔other␈α↔levels.␈α↔ Thus,␈α⊗in␈α↔the␈α↔same␈α↔way␈α↔that␈α↔modifications␈α⊗to
␈↓ α←␈↓information␈αassociated␈αwith␈αthe␈αslotnames␈αcan␈αchange␈αthe␈α
implementation␈αof
␈↓ α←␈↓a␈α
representation␈α
without␈α
impacting␈αits␈α
logical␈α
structure␈α
(exactly␈α
in␈αthe␈α
manner
␈↓ α←␈↓of␈α
record␈α
structures␈α
[Balzer67]),␈α
so␈α
changes␈α
can␈α
be␈α
made␈α
to␈α
logical␈αstructure
␈↓ α←␈↓(the␈α∞schemata)␈α∞without␈α∞impacting␈α
the␈α∞global␈α∞organization␈α∞of␈α
representations
␈↓ α←␈↓␈↓186    KNOWLEDGE ACQUISITION II␈↓ 
∪6-13␈↓

␈↓"β␈↓ α←␈↓(the hierarchy).␈↓
18␈↓
␈↓"β␈↓ α←␈↓␈↓ β?In␈α∪a␈α∪more␈α∩general␈α∪sense,␈α∪both␈α∩stratifications␈α∪provide␈α∪guidance␈α∩in
␈↓ α←␈↓using␈α
the␈αrepresentational␈α
machinery␈αproposed␈α
above.␈α
 In␈αboth␈α
cases␈αwe␈α
have
␈↓ α←␈↓a␈α
set␈α
of␈α
general␈α
guidelines␈αthat␈α
suggests␈α
the␈α
appropriate␈α
mechanism␈α
to␈αuse␈α
for
␈↓ α←␈↓each␈α∂of␈α⊂the␈α∂forms␈α∂of␈α⊂knowledge␈α∂necessary␈α∂for␈α⊂the␈α∂acquisition␈α⊂task.␈α∂ These
␈↓ α←␈↓guidelines␈α∀(Fig.␈α∀6-2␈α∀and␈α∀Fig.␈α∀6-13)␈α∀deal␈α∀with␈α∀dimensions␈α∀of␈α∪knowledge
␈↓ α←␈↓organization␈α
that␈α
are␈αbroadly␈α
applicable␈α
and␈α
hence␈αare␈α
not␈α
limited␈α
to␈αa␈α
single
␈↓ α←␈↓domain␈αof␈αapplication␈αnor␈αto␈αa␈αsingle␈αrepresentational␈αformalism.␈α They␈αthus
␈↓ α←␈↓help to ``make sense'' of the representation scheme outlined here.

␈↓"β␈↓ α←␈↓␈↓α6-14    LIMITATIONS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?One␈α
of␈α
the␈α
primary␈αshortcomings␈α
of␈α
the␈α
current␈α
implementation␈αis␈α
the
␈↓ α←␈↓simplicity␈α⊂of␈α⊂the␈α⊂structure␈α⊂syntax.␈α⊂ While␈α⊂the␈α⊂␈↓↓slotname-blank-advice␈↓␈α⊂triples
␈↓ α←␈↓can␈α∞be␈α∞combined␈α∞in␈α
various␈α∞ways␈α∞and␈α∞the␈α
␈↓↓blank␈↓␈α∞is␈α∞capable␈α∞of␈α∞describing␈α
a
␈↓ α←␈↓range␈α
of␈α∞structures,␈α
the␈α∞result␈α
is␈α
still␈α∞somewhat␈α
rudimentary.␈α∞ The␈α
schemata
␈↓ α←␈↓need␈αa␈αmore␈αpowerful␈αlanguage␈αfor␈αdescribing␈αthe␈α``shape''␈αof␈αdata␈αstructures
␈↓ α←␈↓before they can be widely applicable.
␈↓"β␈↓ α←␈↓␈↓ β?More␈α⊃fundamental␈α⊃limitations␈α⊃arise␈α⊂out␈α⊃of␈α⊃the␈α⊃organization␈α⊃of␈α⊂the
␈↓ α←␈↓slotexperts.␈α They␈αrely,␈αfor␈αinstance,␈αon␈αthe␈αassumption␈αthat␈αknowledge␈αabout
␈↓ α←␈↓the␈α∃representations␈α∃being␈α∃described␈α∃can␈α∃be␈α∃broken␈α∃down␈α⊗into␈α∃basically
␈↓ α←␈↓independent␈α∞chunks␈α∂indexed␈α∞by␈α∂the␈α∞slotname␈α∞and␈α∂advice.␈α∞ This␈α∂requires␈α∞a
␈↓ α←␈↓certain␈α∪modularity␈α∀in␈α∪the␈α∀design␈α∪of␈α∀a␈α∪representation␈α∀that␈α∪is␈α∀not␈α∪always
␈↓ α←␈↓possible to supply.
␈↓"β␈↓ α←␈↓␈↓ β?Independence␈α⊗of␈α∃slotnames␈α⊗implies␈α⊗that␈α∃a␈α⊗representation␈α⊗can␈α∃be
␈↓ α←␈↓decomposed␈αinto␈αa␈αcollection␈αof␈αindependent␈αsubparts,␈αand␈αthis␈αis␈αnot␈αalways
␈↓ α←␈↓true.␈α
 While␈α
the␈αcurrent␈α
implementation␈α
is␈α
able␈αto␈α
deal␈α
with␈α
a␈αlimited␈α
amount
␈↓ α←␈↓of␈α⊃interdependence␈α⊃between␈α∩slots,␈α⊃more␈α⊃complex␈α⊃interdependencies␈α∩do␈α⊃not
␈↓ α←␈↓appear␈α∩to␈α⊃be␈α∩accommodated␈α⊃easily.␈α∩ The␈α⊃current␈α∩implementation␈α∩can,␈α⊃for
␈↓ α←␈↓example,␈α⊂make␈α⊂it␈α⊂possible␈α⊂for␈α⊂one␈α∂slot␈α⊂to␈α⊂use␈α⊂the␈α⊂contents␈α⊂of␈α⊂another␈α∂but
␈↓ α←␈↓cannot␈α
deal␈α
with␈α
the␈α
situation␈α
in␈α
which␈α
the␈α
contents␈α
of␈α
one␈α
slot␈α
restrict␈αthe␈α
set
␈↓ α←␈↓of admissible contents of another.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α_inability␈α_to␈α_deal␈α↔with␈α_more␈α_complex␈α_interrelationships␈α↔of
␈↓ α←␈↓representations␈α_is␈α_currently␈α_the␈α_system's␈α_primary␈α→shortcoming.␈α_ Related
␈↓ α←␈↓attempts␈αto␈αformalize␈αsuch␈αinformation␈α
have␈αcome␈αfrom␈αmany␈αdirections␈α
(e.g.,
␈↓ α←␈↓[Spitzen75],␈α⊗[Stonebreaker75],␈α∃[Suzuki76])␈α⊗and␈α∃have␈α⊗encountered␈α∃similar
␈↓ α←␈↓difficulties.␈α⊗ Specifying␈α⊗complex␈α∃integrity␈α⊗constraints␈α⊗is␈α⊗fundamentally␈α∃a
␈↓ α←␈↓problem␈α∃of␈α⊗knowledge␈α∃representation␈α∃and␈α⊗confronts␈α∃many␈α∃of␈α⊗the␈α∃same
␈↓ α←␈↓difficult issues.
␈↓"β␈↓ α←␈↓␈↓ β?Another␈α∂limitation␈α∂arises␈α∂from␈α∂the␈α∞assumption␈α∂in␈α∂the␈α∂design␈α∂of␈α∞the

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[18]␈α∞Changes␈α
at␈α∞one␈α∞level␈α
may␈α∞quite␈α∞possibly␈α
require␈α∞additional␈α∞changes␈α
at
␈↓ α←␈↓that␈α≠same␈α≠level␈α~in␈α≠order␈α≠to␈α~maintain␈α≠consistency␈α≠of␈α≠data␈α~structure
␈↓ α←␈↓specifications␈α∞or␈α
to␈α∞assure␈α
the␈α∞continued␈α
operation␈α∞of␈α
the␈α∞program.␈α∞ But␈α
by
␈↓ α←␈↓organizing␈αthe␈αinformation␈αin␈αthe␈αlevels␈αdescribed,␈αthe␈αeffects␈αof␈αchanges␈αwill
␈↓ α←␈↓not propagate to the other levels of description.
␈↓ α←␈↓␈↓6-14␈↓ λdLIMITATIONS    187␈↓

␈↓"β␈↓ α←␈↓␈↓↓advice␈↓␈αconcept␈αthat␈αthe␈αquestion␈αof␈αwhere␈αto␈αget␈αthe␈αinformation␈αto␈αfill␈αa␈αslot
␈↓ α←␈↓can␈αbe␈αbroken␈αdown␈αinto␈αa␈αcollection␈αof␈αcases␈αthat␈αare␈α(␈↓↓i␈↓)␈αbroadly␈αapplicable
␈↓ α←␈↓and␈α⊂(␈↓↓ii␈↓)␈α∂independent.␈α⊂ The␈α∂issue␈α⊂is␈α∂not␈α⊂so␈α∂much␈α⊂precisely␈α∂which␈α⊂cases␈α∂are
␈↓ α←␈↓chosen,␈α⊃but␈α⊃that␈α⊃some␈α⊃set␈α⊃of␈α⊃them␈α⊃can␈α⊃be␈α⊃assembled␈α⊃that␈α⊃will␈α⊃provide␈α⊃a
␈↓ α←␈↓``language''␈αfor␈α
designating␈αthe␈α
sources␈αof␈α
knowledge␈αused␈α
in␈αcreating␈α
a␈αdata
␈↓ α←␈↓structure.␈αThe␈αrange␈αof␈αapplication␈αof␈αthe␈αset␈αdetermines␈αthe␈αease␈αwith␈α
which
␈↓ α←␈↓the␈α∃whole␈α∀approach␈α∃can␈α∃be␈α∀used.␈α∃If,␈α∃for␈α∀instance,␈α∃there␈α∃are␈α∀important
␈↓ α←␈↓differences␈α
in␈α
the␈α
implications␈α
of␈α
␈↓	CREATEIT␈↓␈α
for␈α
two␈α
different␈αslotexperts,␈α
then
␈↓ α←␈↓whoever␈αconstructs␈αa␈αschema␈αhas␈αto␈αknow␈αthis␈αfact.␈α But␈αthen␈αlittle␈αis␈αgained
␈↓ α←␈↓by␈α≤the␈α≤whole␈α≠approach.␈α≤ It␈α≤becomes␈α≠far␈α≤less␈α≤transparent,␈α≤and␈α≠the
␈↓ α←␈↓slotname/advice␈α⊗indexing␈α∃scheme␈α⊗becomes␈α⊗an␈α∃obscure␈α⊗way␈α⊗of␈α∃invoking
␈↓ α←␈↓particular␈αpieces␈αof␈αcode.␈α More␈αserious␈αproblems␈αwould␈αarise␈αif␈αthe␈αquestion
␈↓ α←␈↓of␈α
where␈α
to␈α
get␈αthe␈α
information␈α
to␈α
fill␈α
the␈αslots␈α
could␈α
not␈α
even␈αbe␈α
decomposed
␈↓ α←␈↓into any set of distinct, independent cases.
␈↓"β␈↓ α←␈↓␈↓ β?Several␈αof␈αthe␈αtraces␈αdemonstrated␈αthat␈αthe␈αacquisition␈αof␈αone␈αkind␈α
of
␈↓ α←␈↓data␈α∪structure␈α∪can␈α∪lead␈α∪to␈α∪acquisition␈α∪of␈α∪another,␈α∪as␈α∪in␈α∪the␈α∪case␈α∪of␈α∪an
␈↓ α←␈↓attribute␈α∂leading␈α∂to␈α∂its␈α∂associated␈α∂values.␈α∞ This␈α∂is␈α∂a␈α∂useful␈α∂feature,␈α∂since␈α∞it
␈↓ α←␈↓means␈α
that␈α
the␈α
system␈α
tends␈α
to␈α
request␈α
coherent␈α
blocks␈α
of␈α
information␈α
from
␈↓ α←␈↓the␈α∃expert.␈α∃ It␈α⊗depends,␈α∃however,␈α∃on␈α∃explicit␈α⊗structural␈α∃interconnections
␈↓ α←␈↓between␈α⊃data␈α⊃types--the␈α⊃attribute␈α⊂leads␈α⊃to␈α⊃acquiring␈α⊃values␈α⊃because␈α⊂those
␈↓ α←␈↓values␈α
are␈α
part␈α
of␈α
its␈α
structure.␈α
 Had␈α
this␈α
not␈α
been␈α
the␈α
case,␈α
the␈α
link␈αwould
␈↓ α←␈↓not␈αhave␈αbeen␈αmade␈αand␈αthe␈αsystem␈αwould␈αhave␈αacquired␈αeach␈αnew␈αvalue␈αas
␈↓ α←␈↓it␈α∂was␈α∂mentioned.␈α∞ This␈α∂means␈α∂that␈α∂the␈α∞design␈α∂of␈α∂the␈α∂representations␈α∞used
␈↓ α←␈↓can have an important impact on the coherence of the acquisition dialogs.
␈↓"β␈↓ α←␈↓␈↓ β?Requesting␈α⊗the␈α∃expert's␈α⊗help␈α∃in␈α⊗descending␈α∃the␈α⊗schema␈α∃network
␈↓ α←␈↓assumes␈α∞that␈α∞the␈α∞display␈α∞of␈α
the␈α∞alternative␈α∞paths␈α∞will␈α∞be␈α∞comprehensible␈α
to
␈↓ α←␈↓him.␈α⊃ This␈α⊂presumes␈α⊃a␈α⊂correspondence␈α⊃between␈α⊂the␈α⊃representations␈α⊃in␈α⊂the
␈↓ α←␈↓program␈α∞and␈α∞objects␈α∂in␈α∞the␈α∞application␈α∞domain.␈α∂ While␈α∞it␈α∞does␈α∂seem␈α∞likely
␈↓ α←␈↓that␈αsuch␈αa␈αcorrespondence␈αwill␈αexist,␈αits␈αabsence␈αwould␈αpresent␈αa␈αsignificant
␈↓ α←␈↓problem for our system.
␈↓"β␈↓ α←␈↓␈↓ β?While␈α
the␈α
schemata␈α
make␈α
possible␈αa␈α
number␈α
of␈α
useful␈α
features,␈αthey
␈↓ α←␈↓are␈α
not␈α
without␈αassociated␈α
costs.␈α
 Most␈α
of␈αthis␈α
cost␈α
tends␈αto␈α
be␈α
borne␈α
by␈αthe
␈↓ α←␈↓system␈α∩designer,␈α∩to␈α⊃the␈α∩benefit␈α∩of␈α⊃the␈α∩expert␈α∩who␈α⊃wants␈α∩to␈α∩augment␈α⊃the
␈↓ α←␈↓knowledge␈αbase.␈α This␈αis␈αbecause␈αthe␈αschemata␈αimpose␈αa␈αcertain␈αdiscipline␈α
on
␈↓ α←␈↓the␈αsystem␈αdesigner,␈αrequiring,␈αin␈αparticular,␈αthat␈αhe␈αview␈αthe␈αrepresentations
␈↓ α←␈↓in␈α∪the␈α∀system␈α∪in␈α∪fairly␈α∀general␈α∪terms␈α∪and␈α∀fit␈α∪them␈α∪into␈α∀the␈α∪framework
␈↓ α←␈↓provided.␈α
 While␈α
there␈α
are␈α
advantages␈α
to␈α
doing␈α
this,␈α
it␈α
may␈α
not␈α
be␈α∞an␈α
easy
␈↓ α←␈↓task.␈α⊂ Especially␈α⊂during␈α⊂the␈α⊂early␈α⊂stages␈α⊂of␈α⊂system␈α⊂design,␈α⊂when␈α⊂numerous
␈↓ α←␈↓changes␈α∂are␈α∂made,␈α∂the␈α∂cost␈α∂may␈α∂outweigh␈α∂its␈α∂advantages.␈α∂ In␈α∂their␈α∂present
␈↓ α←␈↓state␈α∩of␈α∩development,␈α∪then,␈α∩the␈α∩tools␈α∩described␈α∪in␈α∩this␈α∩chapter␈α∪are␈α∩more
␈↓ α←␈↓appropriate␈α∀to␈α∀ongoing␈α∀knowledge␈α∀base␈α∀maintenance␈α∀than␈α∀to␈α∀the␈α∪initial
␈↓ α←␈↓phases of knowledge base construction.
␈↓"β␈↓ α←␈↓␈↓ β?Perhaps␈α∞the␈α∞most␈α∞general␈α∞limitation␈α∞of␈α∞the␈α∞techniques␈α∂outlined␈α∞here
␈↓ α←␈↓concerns␈α
the␈α
conceptual␈α
level␈α∞of␈α
the␈α
system's␈α
task.␈α∞ It␈α
is␈α
no␈α
accident␈α∞that␈α
we
␈↓ α←␈↓have␈α∞emphasized␈α∞at␈α
many␈α∞points␈α∞the␈α
use␈α∞of␈α∞a␈α
high-level␈α∞language␈α∞and␈α
the
␈↓ α←␈↓␈↓188    KNOWLEDGE ACQUISITION II␈↓ 
∪6-14␈↓

␈↓"β␈↓ α←␈↓manipulation␈α∂of␈α∞extended␈α∂data␈α∞structures␈α∂that␈α∞correspond␈α∂to␈α∞objects␈α∂in␈α∞the
␈↓ α←␈↓world␈α
being␈αmodeled.␈α
 There␈αclearly␈α
are␈αtasks␈α
and␈αsystem␈α
designs␈α
for␈αwhich
␈↓ α←␈↓this␈α⊂correspondence␈α⊂cannot␈α⊂be␈α⊂maintained.␈α⊂ However,␈α⊂to␈α⊂the␈α⊂extent␈α⊂that␈α∂a
␈↓ α←␈↓system␈αcan␈αbe␈αviewed␈αsuccessfully␈αin␈αthese␈αhigh-level␈αterms,␈αthe␈αmethodology
␈↓ α←␈↓can␈α⊂be␈α⊂very␈α∂useful.␈α⊂ In␈α⊂general,␈α∂the␈α⊂higher␈α⊂the␈α∂level␈α⊂of␈α⊂the␈α⊂language␈α∂and
␈↓ α←␈↓programming, the more applicable the techniques will be.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α⊗the␈α⊗current␈α⊗system␈α⊗cannot␈α⊗yet␈α⊗acquire␈α⊗new␈α↔objects␈α⊗(i.e.,
␈↓ α←␈↓``objects''␈α∂as␈α⊂in␈α∂the␈α⊂second␈α∂part␈α⊂of␈α∂an␈α⊂attribute-object-value␈α∂triple)␈α⊂or␈α∂new
␈↓ α←␈↓predicate␈αfunctions.␈α Of␈αthese,␈αthe␈αlatter␈αis␈αmore␈αdifficult,␈αsince␈αit␈αis␈αbasically
␈↓ α←␈↓a␈α
problem␈α
in␈αautomatic␈α
programming␈α
and␈α
no␈αattempt␈α
has␈α
been␈α
made␈αto␈α
solve
␈↓ α←␈↓it.␈α Objects␈αpresent␈αa␈αdifferent␈α
challenge,␈αsince␈αthey␈αare␈αrepresented␈α
by␈αsome
␈↓ α←␈↓highly␈αconvoluted␈αdata␈αstructures␈α
(they␈αare␈αdesigned␈αcurrently␈α
for␈αmaximum
␈↓ α←␈↓efficiency,␈αat␈αthe␈αprice␈αof␈αcomprehensibility).␈α Schema␈αsyntax␈αwill␈αhave␈αto␈αbe
␈↓ α←␈↓extended␈αbefore␈α
it␈αis␈α
capable␈αof␈α
describing␈αthem,␈α
but␈αit␈α
is␈αnot␈α
clear␈αwhether
␈↓ α←␈↓this␈αarises␈α
solely␈αfrom␈α
a␈αshortcoming␈αin␈α
the␈αexpressive␈α
power␈αof␈αthe␈α
schemata
␈↓ α←␈↓or␈α∂whether␈α∂the␈α∂convoluted␈α∂design␈α∂of␈α∂objects␈α∂contributes␈α∂to␈α∂the␈α∂problem␈α∞as
␈↓ α←␈↓well.␈α
 It␈αis␈α
not␈αreasonable,␈α
of␈αcourse,␈α
to␈α
design␈αa␈α
language␈αand␈α
then␈αclaim␈α
that
␈↓ α←␈↓anything␈αit␈α
cannot␈αexpress␈α
should␈αnot␈αbe␈α
said.␈α But␈α
every␈αlanguage␈αcarries␈α
its
␈↓ α←␈↓own␈α
perspective,␈α
and␈α
the␈α
schemata␈αstress␈α
the␈α
simplicity␈α
of␈α
design␈α
that␈αarises
␈↓ α←␈↓from␈αdecomposability.␈α
 (As␈αdiscussed␈α
above,␈αthey␈α
currently␈αrely␈α
␈↓↓too␈↓␈αheavily␈α
on
␈↓ α←␈↓this.)␈α∞One␈α∞of␈α∞the␈α∞potential␈α∞long-term␈α∞benefits␈α∞of␈α∞a␈α∂representation␈α∞language,
␈↓ α←␈↓however,␈α∞is␈α∞as␈α∞a␈α∞vehicle␈α∞for␈α∞developing␈α∞and␈α∞formalizing␈α∞principles␈α∞of␈α
good
␈↓ α←␈↓design.␈α∂ Given␈α∞a␈α∂language␈α∞based␈α∂on␈α∞such␈α∂principles,␈α∞it␈α∂might␈α∞then␈α∂be␈α∞said
␈↓ α←␈↓with␈αsome␈α
justification␈αthat␈α
what␈αcould␈α
not␈αeasily␈α
be␈αstated␈α
might␈αprofit␈α
from
␈↓ α←␈↓reconsideration.␈α∪ The␈α∪schemata␈α∀are,␈α∪naturally,␈α∪only␈α∀a␈α∪single␈α∪step␈α∀in␈α∪this
␈↓ α←␈↓direction and much more work is needed.

␈↓"β␈↓ α←␈↓␈↓α6-15    FUTURE WORK␈↓

␈↓"β␈↓ α←␈↓␈↓α6-15-1    Minor extensions␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈αis␈αapparent␈αfrom␈αa␈αnumber␈αof␈αthe␈αtraces,␈αthe␈αsystem's␈α
``depth-first''
␈↓ α←␈↓approach␈α∩to␈α∩acquisition␈α∩can␈α∩be␈α∩difficult␈α∩to␈α∩follow.␈α∩ Despite␈α∩the␈α∩messages
␈↓ α←␈↓printed␈αby␈αthe␈αsystem␈αand␈αthe␈αindication␈αof␈αlevel␈αgiven␈αby␈αthe␈αindentation␈αof
␈↓ α←␈↓the␈α∂dialog,␈α∞it␈α∂is␈α∞not␈α∂always␈α∞easy␈α∂to␈α∞remember␈α∂which␈α∞problem␈α∂the␈α∂system␈α∞is
␈↓ α←␈↓returning␈αto␈αafter␈αit␈αfinishes␈αup␈αwith␈αa␈αsubproblem.␈α As␈αnoted,␈αthis␈αcould␈αbe
␈↓ α←␈↓solved␈α
by␈α
using␈α
a␈α
modified␈αbreadth-first␈α
search,␈α
in␈α
which␈α
the␈αsystem␈α
finished
␈↓ α←␈↓acquiring␈α
all␈α
the␈α∞necessary␈α
components␈α
at␈α
its␈α∞current␈α
level␈α
before␈α∞taking␈α
up
␈↓ α←␈↓any␈α∩subproblems.␈α⊃ In␈α∩acquiring␈α⊃a␈α∩new␈α⊃attribute,␈α∩for␈α⊃example,␈α∩this␈α⊃would
␈↓ α←␈↓mean␈αthat␈αacquisition␈αof␈αthe␈αnew␈αattribute␈αwould␈αbe␈αfinished␈αbefore␈αstarting
␈↓ α←␈↓to deal with its associated values.
␈↓"β␈↓ α←␈↓␈↓ β?It␈αshould␈αalso␈αbe␈αpossible␈αto␈αsuspend␈αthe␈αacquisition␈αtask␈αtemporarily.
␈↓ α←␈↓The␈αexpert␈αcan␈αat␈αtimes␈αfind␈αhimself␈αinvolved␈αin␈αa␈αprotracted␈αdialog␈αthat␈αis
␈↓ α←␈↓not␈α
immediately␈α
relevant␈α
to␈α
the␈α
bug␈α
he␈α
started␈α
to␈α
correct.␈α
 All␈α
the␈α
information
␈↓ α←␈↓requested␈α⊗will␈α∃prove␈α⊗necessary␈α∃eventually,␈α⊗but␈α∃it␈α⊗may␈α∃prove␈α⊗to␈α⊗be␈α∃an
␈↓ α←␈↓unreasonable␈α
distraction␈αto␈α
have␈α
to␈αdeal␈α
with␈α
every␈αdetail␈α
before␈αgetting␈α
back
␈↓ α←␈↓to the original problem.
␈↓ α←␈↓␈↓6-15␈↓ λMFUTURE WORK    189␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂system␈α∂is␈α∞currently␈α∂an␈α∂effective␈α∞listener␈α∂for␈α∂anyone␈α∂who␈α∞knows
␈↓ α←␈↓just␈α
what␈α
he␈α∞wants␈α
to␈α
say,␈α
but␈α∞it␈α
is␈α
not␈α
at␈α∞all␈α
forgiving.␈α
 It␈α
enforces␈α∞a␈α
one-
␈↓ α←␈↓pass,␈α≠``get-it-all-right-the-first-time''␈α≠approach,␈α≠and␈α≠this␈α≠is␈α≠clearly␈α≠an
␈↓ α←␈↓unrealistic␈α∞view␈α∂of␈α∞knowledge␈α∞base␈α∂development.␈α∞ For␈α∞example,␈α∂the␈α∞schema
␈↓ α←␈↓network␈α
may␈αrequire␈α
reorganization␈αas␈α
a␈αresult␈α
of␈αseveral␈α
causes.␈α
 This␈αmay
␈↓ α←␈↓become␈α
necessary␈α∞because␈α
of␈α∞mistakes␈α
in␈α∞describing␈α
the␈α∞schemata␈α
originally,
␈↓ α←␈↓because␈α
further␈α
development␈α
of␈α
the␈α
performance␈α
program␈α
dictates␈α
redesign␈α
of
␈↓ α←␈↓some␈α
representations,␈α
or␈α
because␈α
the␈α
addition␈α
of␈α
a␈α
new␈α
schema␈α
to␈αthe␈α
network
␈↓ α←␈↓requires␈α∞it␈α∞in␈α
order␈α∞to␈α∞maintain␈α∞the␈α
proper␈α∞inheritance␈α∞of␈α∞properties.␈α
 This
␈↓ α←␈↓makes␈α
the␈α
network␈α
editor␈α
a␈αgood␈α
candidate␈α
for␈α
additional␈α
work.␈α As␈α
currently
␈↓ α←␈↓implemented,␈α∪it␈α∪does␈α∪not␈α∪offer␈α∪any␈α∪mechanism␈α∪for␈α∀reorganizing␈α∪existing
␈↓ α←␈↓schemata;␈α⊂to␈α⊂be␈α⊂a␈α⊂truly␈α⊂useful␈α⊂maintenance␈α⊂tool,␈α⊂it␈α⊂should␈α⊂be␈α⊂extended␈α∂to
␈↓ α←␈↓provide␈α⊃a␈α⊃wider␈α⊃range␈α⊃of␈α⊃such␈α⊃capabilities.␈α⊃ (See␈α⊃[Sandewall75]␈α∩for␈α⊃some
␈↓ α←␈↓suggestions on similar data base reorganization problems.)
␈↓"β␈↓ α←␈↓␈↓ β?Once␈α∞it␈α
becomes␈α∞possible␈α
to␈α∞modify␈α
existing␈α∞representations,␈α∞there␈α
is
␈↓ α←␈↓an␈α
auxiliary␈α
capability␈α
that␈α
would␈α
prove␈α
extremely␈α
useful.␈α
 After␈α
the␈α
user␈α
has
␈↓ α←␈↓finished␈α⊂modifying␈α⊂any␈α⊂schema,␈α⊃the␈α⊂editor␈α⊂should␈α⊂be␈α⊂prepared␈α⊃to␈α⊂execute
␈↓ α←␈↓those␈αsame␈αmodifications␈αon␈αall␈αcurrent␈αinstances␈αof␈αthe␈αschema.␈α That␈αis,␈α
the
␈↓ α←␈↓system␈α∪should␈α∪``look␈α∪over␈α∪the␈α∪user's␈α∪shoulder''␈α∪and␈α∪then␈α∪make␈α∪the␈α∪same
␈↓ α←␈↓changes␈α∞to␈α∞all␈α∞instances␈α∞of␈α∞the␈α∞schema.␈α∞ Simple␈α∞deletions␈α∞or␈α
reorganizations
␈↓ α←␈↓could␈α∂be␈α∂performed␈α∂unaided;␈α∂where␈α∂new␈α∂components␈α∂had␈α∂been␈α⊂added,␈α∂the
␈↓ α←␈↓system␈α∩would␈α∪prompt␈α∩for␈α∪the␈α∩appropriate␈α∩entry␈α∪for␈α∩each␈α∪instance.␈α∩ This
␈↓ α←␈↓would␈α∞allow␈α
extensive␈α∞changes␈α
in␈α∞representation␈α
design␈α∞with␈α∞relatively␈α
little
␈↓ α←␈↓effort and a reduced probability of introducing errors.

␈↓"β␈↓ α←␈↓␈↓α6-15-2    Major extensions␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Perhaps␈αthe␈αmost␈αinteresting␈αmajor␈αextension␈αto␈αthe␈αsystem␈α
would␈αbe
␈↓ α←␈↓the␈α∞addition␈α
of␈α∞semantic␈α
information␈α∞to␈α
the␈α∞schemata.␈α
 They␈α∞were␈α
designed
␈↓ α←␈↓originally␈α∪to␈α∪convey␈α∩the␈α∪syntax␈α∪of␈α∪data␈α∩structures,␈α∪but␈α∪as␈α∪Section␈α∩6-9-2
␈↓ α←␈↓illustrated,␈α⊃inclusion␈α∩of␈α⊃semantic␈α⊃information␈α∩would␈α⊃prove␈α∩very␈α⊃useful--it
␈↓ α←␈↓would␈αmake␈αthe␈αsystem␈αappear␈α``smarter''␈αby␈αallowing␈αit␈αto␈αtake␈αadvantage␈αof
␈↓ α←␈↓context␈α⊃from␈α⊃the␈α⊃debugging␈α⊃dialog␈α⊃to␈α⊃guide␈α⊃its␈α⊃own␈α⊃descent␈α⊃through␈α⊃the
␈↓ α←␈↓schema␈αnetwork.␈α Representation␈αof␈αthe␈αsemantics␈αmight␈αbe␈αbased␈αinitially␈α
on
␈↓ α←␈↓more␈α↔extensive␈α↔use␈α↔of␈α_patterns␈α↔like␈α↔those␈α↔described␈α↔above,␈α_but␈α↔more
␈↓ α←␈↓sophisticated mechanisms should eventually be devised.

␈↓"β␈↓ α←␈↓␈↓α6-16    SUMMARY␈↓

␈↓"β␈↓ α←␈↓␈↓α6-16-1    Review of major concepts␈↓
␈↓"β␈↓ α←␈↓␈↓ β?At␈α∪the␈α∪beginning␈α∪of␈α∪this␈α∪chapter␈α∪we␈α∪suggested␈α∪that␈α∪it␈α∪would␈α∪be
␈↓ α←␈↓instructive␈α
to␈αconsider␈α
the␈α
terms␈α␈↓↓knowledge␈α
representation,␈α
extended␈αdata␈α
type,
␈↓ α←␈↓↓␈↓and␈↓↓␈α∞data␈α∞structure␈↓␈α∞as␈α∞equivalent,␈α∞to␈α∞see␈α∞what␈α∞might␈α∞be␈α∞learned␈α∞by␈α
viewing
␈↓ α←␈↓each␈α
of␈αthem␈α
in␈αthe␈α
perspective␈αnormally␈α
reserved␈αfor␈α
one␈αof␈α
the␈α
others.␈α A
␈↓ α←␈↓number␈α
of␈α
the␈α
key␈α
ideas␈α
involved␈α
in␈α
the␈α
design␈α
and␈α
use␈α
of␈α
the␈αschemata␈α
were
␈↓ α←␈↓inspired by this mixing of perspectives.
␈↓ α←␈↓␈↓190    KNOWLEDGE ACQUISITION II␈↓ 
∪6-16␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂fundamental␈α∞idea␈α∂of␈α∞␈↓↓a␈α∂base␈α∞of␈α∂knowledge␈α∂about␈α∞representations␈↓,
␈↓ α←␈↓for␈α
instance,␈α
was␈α
suggested␈α
by␈α∞the␈α
view␈α
of␈α
representations␈α
as␈α∞extended␈α
data
␈↓ α←␈↓types␈α
and␈αmotivated␈α
by␈α
the␈αdesire␈α
to␈α
organize␈αand␈α
represent␈αknowledge␈α
about
␈↓ α←␈↓those␈α∂data␈α∞types.␈α∂ This␈α∞led␈α∂to␈α∞the␈α∂idea␈α∞of␈α∂the␈α∞␈↓↓schemata␈α∂as␈α∞a␈α∂language␈α∞and
␈↓ α←␈↓↓mechanism␈α∂for␈α∞describing␈α∂representations␈↓,␈α∂and␈α∞it␈α∂strongly␈α∂influenced␈α∞schema
␈↓ α←␈↓design␈α⊂by␈α⊂indicating␈α⊂what␈α⊂sort␈α⊂of␈α⊂information␈α⊂they␈α⊂ought␈α⊂to␈α⊂contain␈α⊂(e.g.,
␈↓ α←␈↓structure␈α
and␈α
interrelationships).␈α∞ This␈α
view␈α
also␈α
suggested␈α∞the␈α
organization
␈↓ α←␈↓of␈α_that␈α→information␈α_and␈α_led␈α→to␈α_␈↓↓organizing␈α_it␈α→around␈α_representational
␈↓ α←␈↓↓primitives␈↓␈α(e.g.,␈αattribute,␈αobject,␈αvalue,␈αetc.),␈αwhich␈αwere,␈αin␈α
turn,␈α␈↓↓represented
␈↓ α←␈↓↓as␈αprototypes␈↓␈α(the␈αschemata)␈αand␈α␈↓↓instantiated␈αto␈αdrive␈αthe␈αinteractive␈αtransfer
␈↓ α←␈↓↓of expertise process␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?Viewing␈α∀extended␈α∀data␈α∀types␈α∀from␈α∀the␈α∀perspective␈α∃of␈α∀knowledge
␈↓ α←␈↓representations␈α
led␈α
to␈α
incorporating␈α
the␈α
␈↓↓advice␈↓␈α
mechanism␈α
in␈α
those␈αdata␈α
types.
␈↓ α←␈↓This␈αprovided␈αan␈αadditional␈αsource␈αof␈αknowledge␈αabout␈αthose␈αstructures␈αand
␈↓ α←␈↓allowed a ``high-level'' dialog that was coherent to the domain expert.
␈↓"β␈↓ α←␈↓␈↓ β?Blurring␈α!the␈α!distinction␈α!between␈α!data␈α!type␈α!and␈α knowledge
␈↓ α←␈↓representation␈α
offered␈α
an␈α
interesting␈α
consideration␈α
for␈α
knowledge␈α
base␈α
design.
␈↓ α←␈↓To␈α
see␈α
how␈α
this␈αconsideration␈α
arose,␈α
note␈α
that␈αa␈α
subtle␈α
factor␈α
that␈α
added␈αto
␈↓ α←␈↓the␈α
coherence␈α
of␈αthe␈α
acquisition␈α
dialogs␈αearlier␈α
was␈α
the␈α
somewhat␈αfortuitous
␈↓ α←␈↓correspondence␈α∪between␈α∪data␈α∪structures␈α∪and␈α∪domain-specific␈α∪objects␈α∪(e.g.,
␈↓ α←␈↓organisms).␈α This␈α
meant␈αthat␈αthe␈α
acquisition␈αdialog␈αappeared␈α
to␈αthe␈αexpert␈α
to
␈↓ α←␈↓be␈α
phrased␈α∞in␈α
terms␈α∞of␈α
objects␈α∞in␈α
the␈α∞domain,␈α
while␈α∞to␈α
the␈α∞system␈α
it␈α∞was␈α
a
␈↓ α←␈↓straightforward␈α∀manipulation␈α∀of␈α∀data␈α∀structures.␈α∀ Such␈α∀a␈α∪correspondence
␈↓ α←␈↓helps␈αto␈αbridge␈αthe␈αgap␈αin␈αperspectives,␈αand␈αthe␈αpurposeful␈αattempt␈αto␈αinsure
␈↓ α←␈↓its␈αpresence␈αin␈αa␈αsystem␈αcan␈αbe␈αa␈αvery␈αsimple,␈αbut␈αuseful␈αconsideration␈αin␈αthe
␈↓ α←␈↓initial design of a knowledge base.
␈↓"β␈↓ α←␈↓␈↓ β?A␈αsecond␈αset␈αof␈αmajor␈αideas␈αinvolved␈αin␈αthe␈αschemata␈αarose␈αfrom␈αthe
␈↓ α←␈↓notion␈αof␈α
␈↓↓levels␈αof␈αknowledge␈↓␈α
described␈αabove.␈α As␈α
noted,␈αthis␈αstratification␈α
of
␈↓ α←␈↓knowledge␈α⊂provided␈α⊂an␈α∂␈↓↓increased␈α⊂range␈α⊂of␈α∂applicability␈↓␈α⊂for␈α⊂the␈α∂techniques
␈↓ α←␈↓and␈α∂offered␈α∂a␈α∂␈↓↓set␈α∂of␈α⊂guidelines␈α∂for␈α∂organizing␈α∂the␈α∂body␈α∂of␈α⊂knowledge␈α∂about
␈↓ α←␈↓↓representations␈↓.␈α⊂ It␈α⊂also␈α⊂suggested␈α⊂that␈α∂the␈α⊂acquisition␈α⊂of␈α⊂new␈α⊂instances␈α∂be
␈↓ α←␈↓viewed␈α∂as␈α∂a␈α⊂process␈α∂of␈α∂␈↓↓descent␈α∂through␈α⊂the␈α∂schema␈α∂hierarchy␈↓,␈α∂and␈α⊂that␈α∂the
␈↓ α←␈↓acquisition␈α
of␈αnew␈α
kinds␈αof␈α
knowledge␈αrepresentations␈α
be␈αviewed␈α
as␈αa␈α
process
␈↓ α←␈↓of ␈↓↓adding new branches to the hierarchy␈↓.

␈↓"β␈↓ α←␈↓␈↓α6-16-2    Current capabilities␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α→schemata␈α→and␈α→associated␈α→structures␈α→offer␈α→a␈α~language␈α→and
␈↓ α←␈↓framework␈α↔in␈α↔which␈α↔representations␈α↔can␈α↔be␈α↔described.␈α_ This␈α↔language
␈↓ α←␈↓strongly␈α
emphasizes␈α∞making␈α
explicit␈α
the␈α∞many␈α
different␈α
kinds␈α∞of␈α
knowledge
␈↓ α←␈↓about␈α≡representations␈α≡and␈α≡offers␈α≡a␈α≡framework␈α≡for␈α≡organizing␈α≡that
␈↓ α←␈↓information.␈α∪ The␈α∪schema␈α∪hierarchy,␈α∪individual␈α∪schemata,␈α∪and␈α∪slotnames
␈↓ α←␈↓each␈α
support␈α
their␈α
own␈α
variety␈α
of␈α
that␈α
knowledge.␈α
 The␈α
result␈α
can␈α
be␈αa␈α
useful
␈↓ α←␈↓global␈α
overview␈αof␈α
the␈α
organization␈αand␈α
design␈α
of␈αall␈α
the␈α
representations␈αin
␈↓ α←␈↓the system.
␈↓"β␈↓ α←␈↓␈↓ β?For␈αboth␈αthe␈αsystem␈αengineer␈αand␈αthe␈αapplications␈αdomain␈αexpert,␈αthe
␈↓ α←␈↓␈↓6-16␈↓ 	βSUMMARY    191␈↓

␈↓"β␈↓ α←␈↓knowledge␈α⊂acquisition␈α⊃capabilities␈α⊂of␈α⊂the␈α⊃schemata␈α⊂offer␈α⊂a␈α⊃very␈α⊂organized
␈↓ α←␈↓and thorough assistant that can:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?␈↓↓attend to many routine details␈↓,
␈↓ α←␈↓␈↓ β?Some␈α∞of␈α∞these␈α∞are␈α
details␈α∞of␈α∞data␈α∞structure␈α∞management,␈α
and
␈↓ α←␈↓␈↓ β?having␈αthe␈α
system␈αattend␈α
to␈αthem␈α
means␈αthe␈α
expert␈αneed␈α
know
␈↓ α←␈↓␈↓ β?nothing␈α&about␈α%programming.␈α&Others␈α%are␈α&details␈α%of
␈↓ α←␈↓␈↓ β?organization␈α∞and␈α∞format,␈α∞and␈α∞with␈α∞these␈α∞out␈α∞of␈α∞the␈α∂way,␈α∞the
␈↓ α←␈↓␈↓ β?task␈αof␈αspecifying␈αlarge␈α
amounts␈αof␈αknowledge␈αbecomes␈αa␈α
good
␈↓ α←␈↓␈↓ β?deal␈α
easier.␈α
 The␈α
emphasis␈α
can␈α
then␈α
be␈α
placed␈α
on␈α
specifying␈α
its
␈↓ α←␈↓␈↓ β?content rather than attending to details of format.

␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?␈↓↓show how knowledge should be specified, and␈↓
␈↓ α←␈↓␈↓ β?In␈α∂terms␈α∂of␈α∂the␈α⊂three␈α∂systems␈α∂pictured␈α∂earlier,␈α⊂the␈α∂assistant's
␈↓ α←␈↓␈↓ β?intelligence␈α~always␈α~lies␈α→at␈α~the␈α~level␈α→above␈α~that␈α~of␈α→the
␈↓ α←␈↓␈↓ β?knowledge␈α≥being␈α≥specified.␈α≤ While␈α≥it␈α≥cannot␈α≥choose␈α≤a
␈↓ α←␈↓␈↓ β?representation␈α↔for␈α↔an␈α↔organism,␈α↔it␈α↔can␈α↔indicate␈α↔how␈α⊗the
␈↓ α←␈↓␈↓ β?representation␈αshould␈αbe␈αspecified.␈α Similarly␈αit␈αcannot␈αsuggest
␈↓ α←␈↓␈↓ β?what␈α∂the␈α∞gramstain␈α∂of␈α∞a␈α∂new␈α∞organism␈α∂might␈α∞be,␈α∂but␈α∂it␈α∞can
␈↓ α←␈↓␈↓ β?indicate␈αthat␈αevery␈αorganism␈αmust␈αhave␈αone,␈αand␈αcan␈αdescribe
␈↓ α←␈↓␈↓ β?exactly␈α⊗how␈α⊗it␈α⊗should␈α⊗be␈α⊗specified.␈α⊗ It␈α⊗is␈α⊗this␈α⊗ability␈α⊗to
␈↓ α←␈↓␈↓ β?structure␈α∂the␈α∂task␈α∂and␈α∂lead␈α∂the␈α∂user␈α∂through␈α∂it␈α∂that␈α⊂is␈α∂most
␈↓ α←␈↓␈↓ β?useful.

␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?␈↓↓make␈α
sure␈α
that␈α
the␈αuser␈α
is␈α
reminded␈α
of␈αall␈α
the␈α
items␈α
he␈α
has␈αto
␈↓ α←␈↓↓␈↓ β?supply␈↓.
␈↓ α←␈↓␈↓ β?Since␈α∂knowledge␈α∂base␈α∂construction␈α∂is␈α∂viewed␈α∂as␈α∂a␈α∂process␈α∞of
␈↓ α←␈↓␈↓ β?knowledge␈α⊃transfer,␈α∩the␈α⊃assistant's␈α⊃thoroughness␈α∩offers␈α⊃some
␈↓ α←␈↓␈↓ β?assurance␈α
that␈α
the␈αtransfer␈α
operations␈α
will␈α
not␈αinadvertently␈α
be
␈↓ α←␈↓␈↓ β?left incomplete.

␈↓"β␈↓ α←␈↓␈↓ β?In␈α∂summary,␈α∂the␈α⊂assistant␈α∂cannot␈α∂supply␈α⊂answers,␈α∂but␈α∂it␈α⊂does␈α∂know
␈↓ α←␈↓what␈α∞all␈α∞the␈α
proper␈α∞questions␈α∞are␈α∞and␈α
what␈α∞constitutes␈α∞a␈α∞syntactically␈α
valid
␈↓ α←␈↓answer␈αfor␈αeach.␈α The␈αapplication␈αdomain␈αexpert␈αwill␈αrely␈αon␈αthe␈αassistant␈αto
␈↓ α←␈↓show␈α∞him␈α∞how␈α∞to␈α∞transfer␈α∂his␈α∞knowledge␈α∞to␈α∞the␈α∞program,␈α∞while␈α∂the␈α∞system
␈↓ α←␈↓designer␈α⊃can␈α⊃use␈α⊃the␈α⊃assistant␈α⊂as␈α⊃an␈α⊃aid␈α⊃in␈α⊃knowledge␈α⊃base␈α⊂management,
␈↓ α←␈↓using␈α
it␈α
to␈α
help␈α
him␈α
keep␈α
track␈α
of␈α
the␈α
large␈α
number␈α
of␈α
representations␈αthat
␈↓ α←␈↓may accumulate during the construction of any sizable program.
␈↓"β␈↓ α←␈↓␈↓ β?All␈α∃of␈α∃this␈α∃should␈α∃make␈α∃plausible␈α∃the␈α∃suggestion␈α∃that␈α∃the␈α∀tools
␈↓ α←␈↓discussed␈α∃above,␈α∃when␈α∀combined␈α∃with␈α∃a␈α∀simple␈α∃core␈α∃of␈α∀representation-
␈↓ α←␈↓independent␈α
information,␈α
offer␈αa␈α
basis␈α
for␈αassembling␈α
a␈α
sizable␈α
collection␈αof
␈↓ α←␈↓knowledge.␈α
 They␈α
provide,␈α∞as␈α
well,␈α
a␈α∞useful␈α
perspective␈α
on␈α∞the␈α
organization
␈↓ α←␈↓and␈α∃representation␈α∃of␈α∃several␈α∃levels␈α∃of␈α∃knowledge,␈α∃making␈α∃the␈α∃transfer
␈↓ α←␈↓process straightforward and effective.
␈↓ α←␈↓␈↓␈↓ 
α    193␈↓




␈↓"β␈↓ α←␈↓␈↓ ε∃␈↓αChapter 7



␈↓"β␈↓ α←␈↓α␈↓ ¬ε␈↓λSTRATEGIES␈↓α


␈↓"β␈↓ α←␈↓α␈↓ ¬7meta-rules for guidance









␈↓"β␈↓ α←␈↓␈↓ ¬¬Know that I have gone many ways wandering in thought. 
␈↓"β␈↓ α←␈↓␈↓ λk␈↓↓Oedipus the King␈↓
␈↓"β␈↓ α←␈↓␈↓ β?06667

␈↓"β␈↓ α←␈↓␈↓α7-1    INTRODUCTION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?In␈α_most␈α↔existing␈α_performance␈α↔programs,␈α_the␈α↔largest␈α_number␈α↔of
␈↓ α←␈↓knowledge␈α
sources␈αrelevant␈α
to␈α
any␈αone␈α
goal␈α
is␈αsmall␈α
enough␈α
that␈αexhaustive
␈↓ α←␈↓invocation␈αis␈αstill␈αcomputationally␈αfeasible.␈↓
1␈↓␈α ␈αIn␈αcurrent␈α(June␈α1978)␈αversions
␈↓ α←␈↓of␈α
␈↓¬MYCIN␈↓,␈α
for␈α
example,␈αthe␈α
largest␈α
number␈α
of␈αrules␈α
relevant␈α
to␈α
a␈αparticular␈α
goal
␈↓ α←␈↓is␈αon␈α
the␈αorder␈αof␈α
50,␈αand␈α
all␈αare␈αinvoked.␈α
 It␈αseems␈α
clear,␈αhowever,␈αthat␈α
since
␈↓ α←␈↓the␈α⊃knowledge␈α⊃bases␈α⊃in␈α⊃such␈α⊃performance␈α⊃programs␈α⊃may␈α⊃eventually␈α⊃grow
␈↓ α←␈↓quite␈α⊂large,␈α⊃exhaustive␈α⊂invocation␈α⊂will␈α⊃in␈α⊂time␈α⊂prove␈α⊃too␈α⊂slow.␈α⊂ It␈α⊃was␈α⊂in
␈↓ α←␈↓response␈α_to␈α_this␈α_that␈α_meta-rules␈α_were␈α_developed,␈α_to␈α→embody␈α_strategic
␈↓ α←␈↓knowledge␈α∩about␈α∪reasoning␈α∩and␈α∪to␈α∩supply␈α∩a␈α∪mechanism␈α∩for␈α∪guiding␈α∩the
␈↓ α←␈↓reasoning process.
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[1]␈α∃Any␈α∃piece␈α∃of␈α∃the␈α∃system␈α∃that␈α∃contributes␈α∃some␈α∃task-specific␈α∃bit␈α∃of
␈↓ α←␈↓intelligence␈αwill␈αbe␈αreferred␈αto␈αas␈αa␈αknowledge␈αsource␈α(KS).␈αThat␈αintelligence
␈↓ α←␈↓may␈αbe␈αat␈αthe␈αlevel␈αof␈αthe␈αprogramming␈αlanguage␈α(e.g.,␈α␈↓¬LISP␈↓␈αfunctions)␈αmight
␈↓ α←␈↓be␈α∞some␈α∂piece␈α∞of␈α∂domain-specific␈α∞information,␈α∞or␈α∂information␈α∞at␈α∂any␈α∞other
␈↓ α←␈↓level␈αon␈αwhich␈αwe␈αcare␈αto␈αfocus.␈αThe␈αprimary␈αconcern␈αhere␈αwill␈αbe␈αwith␈αKSs
␈↓ α←␈↓at␈α⊃the␈α⊃level␈α⊃of␈α⊃problem␈α∩solving;␈α⊃in␈α⊃particular,␈α⊃inference␈α⊃rules␈α⊃of␈α∩the␈α⊃sort
␈↓ α←␈↓shown␈αin␈αFig.␈α2-4.␈α As␈αis␈αcommon,␈αthe␈α``bit''␈αof␈αintelligence␈αwill␈αbe␈αreferred␈αto
␈↓ α←␈↓as␈α∂a␈α∞␈↓↓chunk␈↓.␈α∂ (Since␈α∂strategies␈α∞contribute␈α∂intelligence␈α∂to␈α∞the␈α∂system,␈α∂they␈α∞are
␈↓ α←␈↓KSs␈α∞as␈α∞well.␈α∂For␈α∞the␈α∞sake␈α∞of␈α∂clarity,␈α∞however,␈α∞we␈α∞use␈α∂the␈α∞term␈α∞only␈α∂in␈α∞the
␈↓ α←␈↓context of object-level KSs.)
␈↓ α←␈↓␈↓194    STRATEGIES␈↓ 
#7-1␈↓

␈↓"β␈↓ α←␈↓␈↓ β?Meta-rules␈α∞are␈α∞the␈α∞third␈α∞major␈α
form␈α∞of␈α∞meta-level␈α∞knowledge␈α∞to␈α
be
␈↓ α←␈↓considered.␈↓
2␈↓␈α
They␈α
provide␈α
a␈α
way␈α
of␈α
expressing␈α
knowledge␈α
about␈α
the␈α∞use␈α
of
␈↓ α←␈↓knowledge␈α≥and␈α≤are␈α≥discussed␈α≤here␈α≥as␈α≤a␈α≥framework␈α≥for␈α≤knowledge
␈↓ α←␈↓organization.␈α
 This␈α
chapter␈α
is␈α
divided␈α
into␈α
four␈α
main␈α
parts.␈α
The␈αfirst,␈α
Section
␈↓ α←␈↓7-3,␈α∪gives␈α∪a␈α∪fairly␈α∪general␈α∩introduction␈α∪to␈α∪the␈α∪concept␈α∪of␈α∪strategies␈α∩and
␈↓ α←␈↓considers␈α
their␈α
place␈α
in␈αguiding␈α
program␈α
operation.␈α
 The␈α
second␈αpart,␈α
Section
␈↓ α←␈↓7-4,␈α~discusses␈α~the␈α~specifics␈α~of␈α~meta-rule␈α~structure␈α~and␈α~function␈α→and
␈↓ α←␈↓examines␈α∩their␈α∩contribution␈α∪to␈α∩problem␈α∩solving␈α∩performance.␈α∪ The␈α∩third,
␈↓ α←␈↓Section␈α→7-5,␈α_explores␈α→possible␈α_implications␈α→that␈α_the␈α→current␈α_meta-rule
␈↓ α←␈↓implementation␈α⊃might␈α∩have␈α⊃on␈α⊃programming␈α∩in␈α⊃general.␈α⊃ The␈α∩final␈α⊃part,
␈↓ α←␈↓Section␈α∀7-6,␈α∪then␈α∀offers␈α∀a␈α∪very␈α∀general␈α∀view␈α∪of␈α∀strategies,␈α∀describing␈α∪a
␈↓ α←␈↓taxonomy␈α∃of␈α∃strategy␈α⊗types␈α∃and␈α∃discussing␈α∃the␈α⊗impact␈α∃of␈α∃each␈α⊗on␈α∃the
␈↓ α←␈↓organization of knowledge in a program.

␈↓"β␈↓ α←␈↓␈↓α7-2    THE MAIN IDEAS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Much␈α
of␈α
the␈α
utility␈α
and␈α
impact␈α
of␈α
meta-rules␈α
derives␈α
from␈α
a␈α
few␈α
basic
␈↓ α←␈↓ideas.␈α∞ These␈α∞are␈α∂listed␈α∞and␈α∞described␈α∞below,␈α∂then␈α∞developed␈α∞more␈α∂fully␈α∞in
␈↓ α←␈↓the remainder of this chapter.

␈↓"β␈↓ α←␈↓␈↓ β'Almost␈α∨all␈α∨current␈α∨problem-solving␈α∨control␈α structures␈α∨are
␈↓ α←␈↓␈↓ β'susceptible␈α≠to␈α≠␈↓↓saturation␈↓,␈α≠the␈α≠situation␈α≠in␈α≠which␈α≤so␈α≠many
␈↓ α←␈↓␈↓ β'applicable␈αknowledge␈αsources␈αare␈αretrieved␈αthat␈αit␈αis␈αunrealistic␈αto
␈↓ α←␈↓␈↓ β'consider nonselective, exhaustive invocation.

␈↓ α←␈↓The␈α∂issue␈α∂of␈α∞invocation,␈α∂considered␈α∂in␈α∞the␈α∂most␈α∂general␈α∞terms,␈α∂is␈α∂a␈α∞central
␈↓ α←␈↓theme␈αfor␈αmuch␈αof␈αthe␈αwork␈αdescribed␈αin␈αthis␈αchapter.␈α Over␈αthe␈αyears␈αmany
␈↓ α←␈↓different␈αapproaches␈α
to␈αinvocation␈αhave␈α
been␈αdeveloped,␈α
including␈αstandard
␈↓ α←␈↓subroutine␈α
calling,␈α
varieties␈α
of␈α
pattern-directed␈α
invocation,␈α
etc.␈α
 An␈α
important
␈↓ α←␈↓characteristic␈α⊂of␈α⊂all␈α⊂but␈α⊂the␈α⊂most␈α⊂basic␈α⊂scheme␈α⊂(subroutine␈α⊂calling)␈α⊂is␈α⊂that
␈↓ α←␈↓more␈α∀than␈α∀one␈α∀knowledge␈α∀source␈α∪may␈α∀be␈α∀retrieved␈α∀for␈α∀invocation.␈α∪ All
␈↓ α←␈↓invocation␈αschemes␈α
with␈αthis␈α
property␈αshare␈α
an␈αimportant␈αpotential␈α
weakness,
␈↓ α←␈↓which␈α⊗we␈α∃have␈α⊗labeled␈α⊗␈↓↓saturation␈↓.␈α∃ That␈α⊗is,␈α∃given␈α⊗a␈α⊗sufficiently␈α∃large
␈↓ α←␈↓knowledge␈αbase,␈α
so␈αmany␈αknowledge␈α
sources␈αmay␈αbe␈α
retrieved␈αthat␈αit␈α
becomes
␈↓ α←␈↓impossible␈α⊂to␈α⊃invoke␈α⊂them␈α⊂all.␈α⊃ In␈α⊂that␈α⊂case␈α⊃some␈α⊂decision␈α⊂must␈α⊃be␈α⊂made
␈↓ α←␈↓about how to order and choose from that set.
␈↓"β␈↓ α←␈↓␈↓ β?It␈α_was␈α↔this␈α_issue␈α↔that␈α_motivated␈α↔the␈α_development␈α_of␈α↔strategies
␈↓ α←␈↓described␈αhere,␈α
and␈αit␈αis␈α
this␈αissue␈α
of␈αstrategies␈αas␈α
a␈αsource␈α
of␈αguidance␈αin␈α
the
␈↓ α←␈↓face␈α∞of␈α∞control␈α
structure␈α∞saturation␈α∞that␈α
will␈α∞provide␈α∞much␈α
of␈α∞the␈α∞focus␈α
for
␈↓ α←␈↓this chapter.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α
also␈αconsider␈α
some␈α
extensions␈αof␈α
this␈α
view␈αand␈α
describe,␈αfirst,␈α
how
␈↓ α←␈↓strategies␈α_can␈α_provide␈α_useful␈α_guidance␈α_even␈α_when␈α_a␈α_program␈α_is␈α_not
␈↓ α←␈↓immobilized␈α↔by␈α↔saturation.␈α↔ In␈α_this␈α↔case,␈α↔strategic␈α↔knowledge␈α_allows␈α↔a
␈↓ α←␈↓program␈α_to␈α_proceed␈α→``more␈α_logically,''␈α_without␈α_necessarily␈α→affecting␈α_the

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[2] The others were the rule models and data structure schemata.
␈↓ α←␈↓␈↓7-2␈↓ λ7THE MAIN IDEAS    195␈↓

␈↓"β␈↓ α←␈↓program's␈α
final␈α
result.␈α Second,␈α
we␈α
show␈α
how␈αthe␈α
mechanism␈α
used␈α
to␈αdefine
␈↓ α←␈↓strategies␈α∞has␈α∞interesting␈α∞consequences,␈α∞in␈α∞terms␈α∞of␈α∞allowing␈α∞more␈α∞than␈α
just
␈↓ α←␈↓invocation control; it allows a definition of new invocation criteria as well.

␈↓"β␈↓ α←␈↓␈↓ β'We␈α∂define␈α∂a␈α∂strategy␈α∂to␈α∂be␈α∂information␈α∂about␈α∂which␈α∞knowledge
␈↓ α←␈↓␈↓ β'source(s) to invoke when more than one is potentially useful.

␈↓ α←␈↓Strategies␈α≡are␈α≥thus␈α≡one␈α≡form␈α≥of␈α≡meta-level␈α≡knowledge;␈α≥specifically,
␈↓ α←␈↓information␈α
about␈α
the␈αuse␈α
of␈α
object-level␈αknowledge.␈α
 This␈α
view␈α
provides␈αa
␈↓ α←␈↓useful␈α∞perspective␈α∞on␈α∞the␈α∞organization␈α∞and␈α∞use␈α∞of␈α∞knowledge␈α∞in␈α∞a␈α
program
␈↓ α←␈↓and␈α↔suggests␈α_an␈α↔important␈α↔site␈α_for␈α↔embedding␈α↔knowledge␈α_to␈α↔improve
␈↓ α←␈↓performance.␈α
 We␈αalso␈α
show␈αhow␈α
this␈αview␈α
can␈αbe␈α
generalized␈α
and␈αconsider
␈↓ α←␈↓strategies␈α
as␈α
any␈α
information␈α
about␈α
how␈α
or␈α
when␈α
to␈α
apply␈α
the␈αvarious␈α
sources
␈↓ α←␈↓of object-level knowledge in a program.

␈↓"β␈↓ α←␈↓␈↓ β'Strategies can control invocation by ``tuning'' a control structure.

␈↓ α←␈↓Our␈α∩initial␈α⊃application␈α∩of␈α⊃strategies␈α∩is␈α∩in␈α⊃the␈α∩context␈α⊃of␈α∩a␈α∩single␈α⊃control
␈↓ α←␈↓structure␈α∂and␈α∂shows␈α∂how␈α∂to␈α∂deal␈α∂with␈α∂potential␈α∂saturation␈α∂by␈α∂``tuning''␈α∂the
␈↓ α←␈↓existing knowledge source invocation scheme.

␈↓"β␈↓ α←␈↓␈↓ β'Meta-rules␈αare␈α
implemented␈αby␈α
a␈αtechnique␈αcalled␈α
␈↓↓content-directed
␈↓ α←␈↓↓␈↓ β'invocation␈↓,␈α∃which␈α∃has␈α∀interesting␈α∃implications␈α∃as␈α∃a␈α∀knowledge
␈↓ α←␈↓␈↓ β'source retrieval mechanism.

␈↓ α←␈↓By␈α
␈↓↓content-directed␈α
invocation␈↓,␈αwe␈α
mean␈α
that␈αmeta-rules␈α
refer␈α
to␈αobject-level
␈↓ α←␈↓rules␈α⊃by␈α⊃direct␈α⊃examination␈α⊃of␈α∩the␈α⊃content␈α⊃of␈α⊃the␈α⊃object-level␈α∩rules.␈α⊃ We
␈↓ α←␈↓compare␈α∂this␈α∂to␈α∂more␈α∂traditional␈α∂retrieval␈α∂mechanisms␈α∂and␈α∂demonstrate␈α∂its
␈↓ α←␈↓advantages with respect to ease of modification of the program.

␈↓"β␈↓ α←␈↓␈↓ β'Problem-solving␈α↔control␈α⊗structures␈α↔can␈α↔be␈α⊗viewed␈α↔as␈α↔use␈α⊗of
␈↓ α←␈↓␈↓ β'different retrieval criteria.

␈↓ α←␈↓For␈αexample,␈α
goal-directed␈αinvocation␈α
involves␈αretrieving␈α
knowledge␈αsources
␈↓ α←␈↓by␈α
the␈αgoal␈α
they␈αaccomplish;␈α
data-directed␈αinvocation␈α
selects␈αon␈α
the␈α
basis␈αof
␈↓ α←␈↓the␈α∀data␈α∀available;␈α∀while␈α∀means-ends␈α∪analysis␈α∀retrieves␈α∀on␈α∀the␈α∀basis␈α∪of
␈↓ α←␈↓differences between the current state and the goal state.

␈↓"β␈↓ α←␈↓␈↓ β'Content-directed␈α≤invocation␈α≠provides␈α≤an␈α≤explicit,␈α≠functional
␈↓ α←␈↓␈↓ β'definition␈α_of␈α→retrieval␈α_criteria␈α_and␈α→offers␈α_a␈α→mechanism␈α_for
␈↓ α←␈↓␈↓ β'defining new criteria.

␈↓ α←␈↓␈↓¬TEIRESIAS␈↓␈α
thus␈αprovides␈α
an␈αenvironment␈α
in␈αwhich␈α
the␈αretrieval␈α
criteria␈αare␈α
not
␈↓ α←␈↓predefined␈α∃and␈α⊗embedded␈α∃in␈α∃the␈α⊗interpreter␈α∃(as␈α∃in␈α⊗most␈α∃programming
␈↓ α←␈↓languages)␈α∞but␈α∞are,␈α∂instead,␈α∞available␈α∞to␈α∞the␈α∂programmer.␈α∞ We␈α∞will␈α∂see␈α∞this
␈↓ α←␈↓␈↓196    STRATEGIES␈↓ 
#7-2␈↓

␈↓"β␈↓ α←␈↓idea␈α
applied␈αprimarily␈α
in␈αthe␈α
context␈αof␈α
tuning␈αan␈α
existing␈α
control␈αstructure
␈↓ α←␈↓(e.g.,␈αdoing␈α
goal-directed␈αinvocation␈α
that␈αis␈αalso␈α
``speed-directed,''␈αin␈α
that␈αthe
␈↓ α←␈↓relevant␈α∩operators␈α∩are␈α⊃ordered␈α∩on␈α∩the␈α⊃basis␈α∩of␈α∩speed).␈α⊃ The␈α∩idea␈α∩is␈α⊃also
␈↓ α←␈↓considered␈α
briefly␈α
as␈αthe␈α
basis␈α
for␈αa␈α
program␈α
that␈αadaptively␈α
selects␈α
its␈αown
␈↓ α←␈↓control structures.

␈↓"β␈↓ α←␈↓␈↓α7-3    WHAT IS A STRATEGY␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αword␈α``strategy''␈αhas␈αbeen␈αused␈αin␈αmany␈αcontexts␈αand␈αsenses,␈αand
␈↓ α←␈↓seems␈αto␈αbe␈αa␈αsomewhat␈αelusive␈α
concept.␈α We␈αare␈αconcerned␈αhere␈αmainly␈α
with
␈↓ α←␈↓its␈α
impact␈α
in␈α
the␈α
context␈α
of␈α
problem␈α
solving␈α
and␈α
offer␈α
a␈α
definition␈α
in␈α
these
␈↓ α←␈↓terms.␈α To␈αhelp␈αset␈αthe␈αcontext,␈αwe␈αconsider␈αfirst␈αa␈αbroadly␈αconstrued␈αview␈αof
␈↓ α←␈↓problem solving and examine the notion of ill structured problems.

␈↓"β␈↓ α←␈↓␈↓α7-3-1    Ill structured problems␈↓
␈↓"β␈↓ α←␈↓␈↓ β?To␈αcharacterize␈αthe␈αclass␈αof␈αproblems␈αmost␈αrelevant␈αto␈αthe␈αframework
␈↓ α←␈↓presented␈α
below,␈αwe␈α
consider␈α
the␈αdistinction␈α
between␈α
a␈α␈↓↓well␈α
structured␈↓␈αand␈α
an
␈↓ α←␈↓␈↓↓ill structured␈↓ problem.  The former is described in [Newell69] as a problem:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?which can be stated in terms of numerical variables,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?whose␈α∃goals␈α∃can␈α∃be␈α⊗specified␈α∃in␈α∃terms␈α∃of␈α⊗a␈α∃well-defined
␈↓ α←␈↓␈↓ β?objective function, and
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?for which there exists an algorithmic solution.

␈↓"β␈↓ α←␈↓␈↓ β?Ill␈α
structured␈α∞problems␈α
are␈α∞those␈α
which␈α
do␈α∞not␈α
meet␈α∞one␈α
or␈α∞more␈α
of
␈↓ α←␈↓the␈α∀above␈α∀conditions.␈↓
3␈↓␈α∀The␈α∀most␈α∃general␈α∀such␈α∀case␈α∀would␈α∀lack␈α∃even␈α∀a
␈↓ α←␈↓definition␈αof␈αthe␈αinitial␈αstate␈αand␈αa␈αspecification␈αof␈αthe␈αgoal.␈α Most␈αproblems
␈↓ α←␈↓attacked␈αby␈αAI,␈αhowever,␈αhave␈α
well-defined␈αinitial␈αand␈αgoal␈αstates;␈α
the␈αthird
␈↓ α←␈↓condition␈α∂is␈α∂the␈α∞one␈α∂most␈α∂often␈α∞not␈α∂met.␈α∂ Note␈α∞that␈α∂the␈α∂domain␈α∂of␈α∞clinical
␈↓ α←␈↓medicine␈αis␈αmore␈αill␈αstructured␈αthan␈αmost␈α
since␈αin␈αmany␈αcases␈αit␈αis␈αnot␈α
certain
␈↓ α←␈↓what constitutes the ``correct'' diagnosis (and hence therapy).
␈↓"β␈↓ α←␈↓␈↓ β?To␈αillustrate␈αthis␈αdistinction␈α
between␈αwell␈αstructured␈αand␈αill␈α
structured
␈↓ α←␈↓problems,␈α
consider␈α
the␈α
task␈α
of␈α
the␈α
␈↓¬STUDENT␈↓␈α
program␈α
[Bobrow68],␈α
which␈α
dealt
␈↓ α←␈↓with␈αthe␈αdomain␈αof␈αsimple␈αalgebra␈αword␈αproblems.␈α It␈αfirst␈αcreated␈αequations
␈↓ α←␈↓that␈α
corresponded␈α
to␈αthe␈α
natural␈α
language␈αtext,␈α
then␈α
solved␈α
those␈αequations.
␈↓ α←␈↓The␈α∩problem␈α⊃of␈α∩turning␈α⊃the␈α∩natural␈α∩language␈α⊃into␈α∩a␈α⊃set␈α∩of␈α∩equations␈α⊃is
␈↓ α←␈↓highly␈α∀ill␈α∀structured.␈α∀ Once␈α∀the␈α∀equations␈α∀are␈α∀determined,␈α∀however,␈α∀the
␈↓ α←␈↓process of solving them is perfectly straightforward.
␈↓"β␈↓ α←␈↓␈↓ β?For␈α
purposes␈α
of␈α
this␈α
discussion␈αwe␈α
focus␈α
on␈α
the␈α
behavior␈αdisplayed␈α
by
␈↓ α←␈↓programs␈α∩that␈α∩attempt␈α∩to␈α∩solve␈α∩ill␈α∩structured␈α∩problems.␈α∩ We␈α∪examine,␈α∩in

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[3]␈α∂Simon␈α∂has␈α∂argued␈α∂[Simon73]␈α∂that␈α∂the␈α∂division␈α∂between␈α∂well␈α∞structured
␈↓ α←␈↓and␈α∂ill␈α∞structured␈α∂problems␈α∂is␈α∞less␈α∂precise␈α∂and␈α∞that␈α∂even␈α∂problems␈α∞meeting
␈↓ α←␈↓these␈α
criteria␈α
in␈α
principle␈αmay␈α
be␈α
ill␈α
structured␈α
in␈αpractice.␈α
 This␈α
is␈α
due␈αto␈α
the
␈↓ α←␈↓unrealistic␈α⊃amount␈α⊃of␈α⊃computation␈α⊃required␈α⊃to␈α⊃satisfy␈α⊃condition␈α⊃(c)␈α⊃above
␈↓ α←␈↓when a program's knowledge base is very large.
␈↓ α←␈↓␈↓7-3␈↓ π{WHAT IS A STRATEGY    197␈↓

␈↓"β␈↓ α←␈↓particular,␈α
the␈α
degree␈αof␈α
nondeterminacy␈α
in␈αthe␈α
process␈α
of␈αselecting␈α
a␈α
KS␈αto
␈↓ α←␈↓invoke,␈αand␈αshow␈αthat␈αthe␈αnondeterminacy␈αpresent␈αin␈αill␈αstructured␈αproblems
␈↓ α←␈↓is a common source of saturation.
␈↓"β␈↓ α←␈↓␈↓ β?To␈αsee␈α
this,␈αconsider␈αthat␈α
well␈αstructured␈αproblems␈α
(like␈αsolving␈α
a␈αset
␈↓ α←␈↓of␈α∞linear␈α
equations)␈α∞can␈α
(by␈α∞definition)␈α
be␈α∞solved␈α
algorithmically.␈α∞ Hence␈α
at
␈↓ α←␈↓each␈α∃point␈α∃of␈α⊗the␈α∃problem␈α∃solution␈α∃there␈α⊗is␈α∃no␈α∃question␈α⊗about␈α∃which
␈↓ α←␈↓knowledge␈α∞source␈α∞should␈α
be␈α∞invoked␈α∞next␈α∞and␈α
no␈α∞question␈α∞of␈α∞whether␈α
this
␈↓ α←␈↓invocation␈α∞gets␈α∞us␈α∞closer␈α∞to␈α∞the␈α∞solution.␈α∞ Since␈α∞ill␈α∞structured␈α∂problems␈α∞lack
␈↓ α←␈↓algorithmic␈α
solutions,␈α∞the␈α
choice␈α
of␈α∞a␈α
KS␈α
to␈α∞invoke␈α
is␈α
at␈α∞best␈α
a␈α∞good␈α
guess.
␈↓ α←␈↓In such a case, we may be faced with a problem for which

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?it␈α
may␈α
be␈α
useful␈α
to␈α∞make␈α
explicit␈α
in␈α
the␈α
program␈α∞the␈α
process
␈↓ α←␈↓␈↓ β?by␈α∩which␈α⊃a␈α∩KS␈α⊃is␈α∩selected␈α⊃(as␈α∩opposed␈α⊃to␈α∩the␈α∩situation␈α⊃in
␈↓ α←␈↓␈↓ β?algorithmic␈α_programs,␈α↔where␈α_only␈α_the␈α↔end␈α_result␈α_of␈α↔the
␈↓ α←␈↓␈↓ β?selection␈αprocess␈α
is␈αevident,␈αin␈α
the␈αpredefined␈αsequence␈α
of␈αKS
␈↓ α←␈↓␈↓ β?invocations);

␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?the␈αresult␈αof␈α
the␈αselection␈αprocess␈α
may␈αbe␈αa␈α
number␈αof␈αKSs,␈α
all
␈↓ α←␈↓␈↓ β?of␈α∞which␈α∞are␈α∞potentially␈α∂useful␈α∞(as␈α∞opposed␈α∞to␈α∞the␈α∂single␈α∞KS
␈↓ α←␈↓␈↓ β?[pre]selected␈α∪for␈α∩invocation␈α∪at␈α∪each␈α∩step␈α∪of␈α∪an␈α∩algorithmic
␈↓ α←␈↓␈↓ β?program); and

␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?the␈α∀process␈α∀of␈α∪selecting␈α∀a␈α∀KS␈α∪for␈α∀invocation␈α∀needs␈α∀to␈α∪be
␈↓ α←␈↓␈↓ β?carried␈α
out␈α
often␈α
(i.e.,␈α
invocation␈α
of␈α
a␈α
particular␈α
KS␈α
does␈α
not
␈↓ α←␈↓␈↓ β?carry␈α⊂the␈α⊂computation␈α⊂very␈α⊂far,␈α⊂and␈α⊂many␈α⊃such␈α⊂invocations
␈↓ α←␈↓␈↓ β?are needed before reaching a solution).

␈↓"β␈↓ α←␈↓␈↓ β?As␈α∂will␈α∂become␈α⊂clear␈α∂in␈α∂the␈α∂remainder␈α⊂of␈α∂this␈α∂chapter,␈α∂all␈α⊂of␈α∂these
␈↓ α←␈↓were␈α∞important␈α∞considerations␈α
in␈α∞the␈α∞concept␈α
and␈α∞design␈α∞of␈α∞meta-rules.␈α
 At
␈↓ α←␈↓the␈α∂moment,␈α∞however,␈α∂it␈α∂is␈α∞useful␈α∂to␈α∂focus␈α∞on␈α∂the␈α∂second,␈α∞since␈α∂it␈α∂leads␈α∞us
␈↓ α←␈↓back␈αto␈αthe␈αidea␈αof␈αsaturation.␈α That␈αis,␈αone␈αimportant␈αcause␈αof␈αsaturation␈αis
␈↓ α←␈↓the nondeterminism present in ill structured problems.
␈↓"β␈↓ α←␈↓␈↓ β?When␈αill␈αstructured␈αproblems␈α
are␈αattacked␈αwith␈αthe␈αproduction␈α
system
␈↓ α←␈↓methodology,␈αthe␈αconcept␈αof␈α``degree␈αof␈αnondeterminacy''␈αhas␈αa␈αwell-specified
␈↓ α←␈↓instantiation: ␈α
It␈α
is␈α
called␈α
the␈α
␈↓↓conflict␈α
set␈↓.␈α
This␈α
is␈α
the␈α
set␈α
of␈α
all␈α
rules␈α
that,␈αat
␈↓ α←␈↓any␈α↔given␈α↔moment,␈α↔meet␈α↔the␈α↔tests␈α↔for␈α↔applicability␈α↔and,␈α_hence,␈α↔could
␈↓ α←␈↓justifiably␈αbe␈αused.␈αDepending␈αon␈αthe␈αnature␈αof␈αthe␈αproblem,␈αthe␈αconflict␈αset
␈↓ α←␈↓may range from a single rule to every rule in the system.
␈↓"β␈↓ α←␈↓␈↓ β?It␈α∞will␈α∞be␈α∞useful␈α∞to␈α∞generalize␈α∞this␈α∞concept␈α∞slightly␈α∞so␈α∞that␈α∞it␈α∞will␈α∞be
␈↓ α←␈↓applicable␈α↔to␈α↔additional␈α↔programming␈α↔methodologies.␈α↔ Rather␈α↔than␈α↔the
␈↓ α←␈↓conflict␈αset,␈αwe␈αwill␈αspeak␈αof␈αthe␈α␈↓↓plausible␈αknowledge␈αsource␈αset␈↓␈α(PKS␈αset)␈αand
␈↓ α←␈↓mean␈α∞by␈α∞that␈α∞the␈α∞set␈α∞of␈α∞all␈α∞KSs␈α∞that,␈α∞at␈α∞some␈α∞given␈α∞moment,␈α∞are␈α
plausibly
␈↓ α←␈↓useful␈α⊂and␈α⊂appropriate␈α⊂to␈α⊂invoke.␈α∂ This␈α⊂will␈α⊂also␈α⊂serve␈α⊂to␈α⊂emphasize␈α∂that
␈↓ α←␈↓saturation␈αis␈αindependent␈αof␈αany␈αspecific␈αknowledge␈αrepresentation␈αor␈αcontrol
␈↓ α←␈↓structure.␈α Any␈αsystem,␈αfaced␈αwith␈αa␈αsufficiently␈αill␈αstructured␈αproblem␈αand␈αa
␈↓ α←␈↓␈↓198    STRATEGIES␈↓ 
#7-3␈↓

␈↓"β␈↓ α←␈↓large␈α
enough␈αknowledge␈α
base,␈α
may␈αbe␈α
unable␈αto␈α
select␈α
out␈αfew␈α
enough␈αKSs␈α
to
␈↓ α←␈↓make exhaustive invocation feasible.
␈↓"β␈↓ α←␈↓␈↓ β?By␈α∀defining␈α∀two␈α∃averages␈α∀on␈α∀this␈α∀set,␈α∃we␈α∀can␈α∀construct␈α∃a␈α∀more
␈↓ α←␈↓quantitative␈αinterpretation␈αof␈α
the␈αconcept␈αof␈α
well␈αstructured␈αand␈αill␈α
structured
␈↓ α←␈↓problems.␈α⊃ The␈α⊃``average␈α⊃size''␈α⊃of␈α⊃the␈α∩PKS␈α⊃set␈α⊃will␈α⊃be␈α⊃defined␈α⊃as␈α∩a␈α⊃time
␈↓ α←␈↓average␈α∩of␈α∩its␈α∩size␈α∩over␈α∩the␈α⊃course␈α∩of␈α∩the␈α∩entire␈α∩problem␈α∩solution.␈α⊃ The
␈↓ α←␈↓``average␈α
power''␈α
will␈α∞be␈α
defined␈α
as␈α∞the␈α
average␈α
of␈α
the␈α∞power␈α
of␈α
each␈α∞of␈α
its
␈↓ α←␈↓members.␈↓
4␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∃``degree␈α∃of␈α∃ill-structuredness''␈α∀of␈α∃a␈α∃problem␈α∃should␈α∃then␈α∀be
␈↓ α←␈↓directly␈α∀proportional␈α∀to␈α∀the␈α∀average␈α∀size␈α∀of␈α∀the␈α∀PKS␈α∀set␈α∀and␈α∪inversely
␈↓ α←␈↓proportional␈α
to␈α
its␈α
average␈α
power.␈α
 (This␈α
is␈α
of␈α
course␈α
a␈α
relative␈α
rather␈αthan
␈↓ α←␈↓absolute␈α∂scale,␈α∂whose␈α∂utility␈α∂lies␈α∂in␈α∂making␈α∂broad␈α∂comparisons␈α∂rather␈α∂than
␈↓ α←␈↓independent␈αmeasurements.)  ␈αIf␈αthis␈α
number␈αis␈αrelatively␈αsmall,␈α
the␈αproblem
␈↓ α←␈↓is␈α
more␈α
likely␈α
to␈α∞be␈α
well␈α
structured␈α
and␈α
not␈α∞the␈α
type␈α
of␈α
problem␈α
we␈α∞will␈α
be
␈↓ α←␈↓considering␈α⊗here.␈α⊗The␈α⊗larger␈α⊗this␈α⊗number␈α⊗gets,␈α⊗the␈α⊗more␈α↔relevant␈α⊗the
␈↓ α←␈↓remainder of this analysis becomes.

␈↓"β␈↓ α←␈↓␈↓α7-3-2    Strategies␈↓
␈↓"β␈↓ α←␈↓␈↓ β?With␈αthis␈αbackground␈αwe␈αcan␈α
now␈αexamine␈αthe␈αnotion␈αof␈α
strategy␈αin
␈↓ α←␈↓more␈α⊂detail.␈α⊂ We␈α⊂present␈α⊂three␈α⊂successively␈α⊂more␈α⊂general␈α⊃views,␈α⊂beginning
␈↓ α←␈↓with a definition in terms of coping with saturation.

␈↓"β␈↓ α←␈↓␈↓αStrategies as a response to saturation␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Faced␈αwith␈αa␈α
PKS␈αset␈αof␈α
nontrivial␈αsize␈αand␈α
a␈αdesire␈αto␈α
avoid␈αblind,
␈↓ α←␈↓exhaustive␈αinvocation,␈αsome␈α
decision␈αmust␈αbe␈α
made␈αabout␈αwhich␈α
KS␈αshould
␈↓ α←␈↓be␈αthe␈αnext␈αto␈αbe␈αinvoked.␈α It␈αis␈αour␈αcontention␈αthat␈αthis␈αdecision␈αpoint␈αis␈αan
␈↓ α←␈↓important␈αsite␈α
for␈αthe␈αembedding␈α
of␈αknowledge,␈αbecause␈α
system␈αperformance
␈↓ α←␈↓will␈α⊃be␈α⊃strongly␈α⊃influenced␈α⊃by␈α⊂the␈α⊃intelligence␈α⊃with␈α⊃which␈α⊃the␈α⊃decision␈α⊂is
␈↓ α←␈↓made.␈αWe␈αclaim␈αalso␈αthat␈αit␈αis␈αa␈αdecision␈αwhich␈αin␈αmany␈αsystems␈αis␈αmade␈αon
␈↓ α←␈↓the␈α∪basis␈α∪of␈α∀knowledge␈α∪that␈α∪is␈α∪neither␈α∀explicit,␈α∪nor␈α∪well␈α∀organized,␈α∪nor
␈↓ α←␈↓considered in appropriate generality.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α∂take␈α∂this␈α∂as␈α∂a␈α∂starting␈α∂point␈α∂and␈α∂offer␈α∂our␈α∂initial␈α∂definition␈α∂of
␈↓ α←␈↓strategy␈α
in␈α
terms␈αof␈α
it: ␈α
We␈αsuggest␈α
that␈α
a␈αstrategy␈α
can␈α
profitably␈α
be␈αviewed
␈↓ α←␈↓as␈α∂␈↓↓information␈α∞concerning␈α∂which␈α∞chunk␈α∂of␈α∞knowledge␈α∂might␈α∞be␈α∂invoked␈α∞next,
␈↓ α←␈↓↓when more than one chunk may be applicable␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αutility␈αof␈αthis␈αview␈αof␈αstrategies␈αarises␈αfrom␈αits␈αapplicability␈αboth
␈↓ α←␈↓to␈αproblem␈α
solving␈α(in␈αparticular,␈α
for␈αill␈αstructured␈α
problems)␈αand␈α
to␈αcurrent
␈↓ α←␈↓developments␈α∃in␈α∀programming␈α∃languages.␈α∀ Invocation␈α∃in␈α∃traditional␈α∀(i.e.,
␈↓ α←␈↓␈↓¬ALGOL␈↓-like)␈α∞programs,␈α∞for␈α∞example,␈α∞is␈α∞``well␈α∞specified,''␈α∞in␈α∞the␈α∞sense␈α∂that␈α∞only

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[4]␈α∂``Power''␈α∂is␈α∂the␈α∂concept␈α⊂defined␈α∂by␈α∂Newell␈α∂[Newell69]␈α∂as␈α∂``the␈α⊂ability␈α∂to
␈↓ α←␈↓deliver␈α∞solutions.'' ␈α∞He␈α
describes␈α∞it␈α∞as␈α∞composed␈α
of␈α∞(among␈α∞other␈α∞things): ␈α
a
␈↓ α←␈↓KS's␈αprobability␈αof␈αfinding␈α
a␈αsolution,␈αthe␈αquality␈α
of␈αthe␈αsolution␈α(optimal,␈α
or
␈↓ α←␈↓how␈α∂suboptimal),␈α∂and␈α∂the␈α∂computational␈α∂expense␈α∂incurred␈α∂in␈α∂invoking␈α∂the
␈↓ α←␈↓KS.
␈↓ α←␈↓␈↓7-3␈↓ π{WHAT IS A STRATEGY    199␈↓

␈↓"β␈↓ α←␈↓one␈α
procedure␈α
is␈α∞considered␈α
for␈α
invocation␈α
at␈α∞any␈α
given␈α
moment.␈α∞ Many␈α
of
␈↓ α←␈↓the␈α⊃programming␈α⊃paradigms␈α⊃developed␈α⊃more␈α⊃recently,␈α⊃however,␈α⊃admit␈α⊂(or
␈↓ α←␈↓even␈αencourage)␈αthe␈αpossibility␈αof␈αretrieving␈αseveral␈αchunks␈αof␈αknowledge,␈αall
␈↓ α←␈↓of␈α∩which␈α∩are␈α∩plausibly␈α∩useful␈α∩in␈α∩a␈α∩single␈α∩situation␈α∩(e.g.,␈α∩this␈α∩is␈α∪true␈α∩for
␈↓ α←␈↓production␈α∪rules,␈α∪␈↓¬PLANNER␈↓-like␈α∪languages,␈α∀as␈α∪well␈α∪as␈α∪other␈α∀languages␈α∪with
␈↓ α←␈↓choice-point␈α∞and␈α∞backtracking␈α∞mechanisms).␈α∞ Typically,␈α∞in␈α∞these␈α∞paradigms,
␈↓ α←␈↓the␈α↔KSs␈α↔are␈α↔retrieved␈α_unordered␈α↔and␈α↔are␈α↔invoked␈α_exhaustively,␈α↔each
␈↓ α←␈↓considered␈αin␈α
turn,␈αuntil␈α
some␈αstopping␈α
criterion␈αis␈α
met␈αor␈α
until␈αall␈αhave␈α
been
␈↓ α←␈↓tried.␈α⊃ However,␈α⊃faced␈α⊃with␈α⊃a␈α⊃set␈α⊃of␈α⊃alternatives␈α⊃large␈α⊃enough␈α⊃(or␈α⊂varied
␈↓ α←␈↓enough)␈αthat␈αblind,␈αexhaustive␈αinvocation␈αwould␈αbe␈αinfeasible,␈αsome␈αdecision
␈↓ α←␈↓must␈αbe␈α
made␈αabout␈αwhich␈α
should␈αbe␈α
chosen.␈↓
5␈↓␈α ␈αThe␈α
concept␈αof␈αstrategy␈α
thus
␈↓ α←␈↓appears␈α⊂to␈α∂have␈α⊂a␈α∂natural␈α⊂place␈α∂in␈α⊂considering␈α∂questions␈α⊂of␈α∂programming
␈↓ α←␈↓language design, as well as problem solving.

␈↓"β␈↓ α←␈↓␈↓αStrategies as meta-level knowledge␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Considered␈α∞in␈α∞more␈α∞general␈α∞terms,␈α
strategies␈α∞are␈α∞a␈α∞third␈α∞example␈α
of
␈↓ α←␈↓meta-level␈αknowledge,␈αin␈αparticular,␈αknowledge␈αabout␈αhow␈α(and␈αwhen)␈αto␈αuse
␈↓ α←␈↓the various sources of object-level knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?To␈α
illustrate␈α
this␈α
concept␈α
and␈α
emphasize␈α
that␈α
it␈α
is␈α∞widely␈α
applicable,
␈↓ α←␈↓three␈α⊂common␈α∂approaches␈α⊂to␈α∂problem␈α⊂solving␈α∂are␈α⊂listed␈α∂below,␈α⊂along␈α∂with
␈↓ α←␈↓examples␈α∞of␈α∞object-level␈α∞and␈α∞meta-level␈α∞knowledge␈α∞for␈α∞each.␈α∞ In␈α∂each␈α∞case,
␈↓ α←␈↓the␈αmeta-level␈αknowledge␈αconsists␈αof␈αadvice␈αabout␈αhow␈αand␈αwhen␈αto␈α
use␈αthe
␈↓ α←␈↓various sources of object-level knowledge.

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?Problem decomposition:


␈↓"β␈↓ α←␈↓␈↓ β?␈↓↓object␈α⊃level␈↓--knowing␈α⊂how␈α⊃to␈α⊃decompose␈α⊂the␈α⊃problem␈α⊃(i.e.,␈α⊂knowing
␈↓ α←␈↓what the necessary and sufficient subgoals are),

␈↓"β␈↓ α←␈↓␈↓ β?␈↓↓meta␈α∩level␈↓--knowing␈α∩how␈α⊃to␈α∩order␈α∩the␈α⊃attack␈α∩on␈α∩the␈α∩subgoals␈α⊃for
␈↓ α←␈↓efficiency (i.e., knowing which of several possible decompositions to try);







␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[5]␈α
A␈α
few␈α
of␈α
the␈α
paradigms␈α
offer␈α
mechanisms␈α
for␈α
doing␈α
this.␈α
 For␈αexample,
␈↓ α←␈↓the␈α
concept␈α
of␈α
␈↓↓conflict␈α
resolution␈↓␈αused␈α
with␈α
production␈α
rules␈α
offers␈α
a␈αlimited
␈↓ α←␈↓means␈α∞of␈α∂ordering␈α∞the␈α∞KSs␈α∂retrieved,␈α∞but␈α∂its␈α∞use␈α∞thus␈α∂far␈α∞has␈α∂been␈α∞purely
␈↓ α←␈↓syntactic.␈α It␈αmay,␈αfor␈αinstance,␈αorder␈αrules␈αbased␈αon␈αwhich␈αof␈αthose␈αretrieved
␈↓ α←␈↓was␈α_used␈α_least␈α_recently.␈α_ As␈α_we␈α_will␈α_see,␈α_␈↓¬PLANNER␈↓'s␈α_␈↓	THUSE␈↓␈α→and␈α_␈↓	THTBF␈↓
␈↓ α←␈↓mechanisms are somewhat more sophisticated.
␈↓ α←␈↓␈↓200    STRATEGIES␈↓ 
#7-3␈↓


␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?Cooperating knowledge sources:


␈↓"β␈↓ α←␈↓␈↓ β?␈↓↓object level␈↓--the domain knowledge contained in the various sources,

␈↓"β␈↓ α←␈↓␈↓ β?␈↓↓meta␈α
level␈↓--knowing␈αwhich␈α
knowledge␈α
source␈αto␈α
use␈αat␈α
each␈α
point␈αin
␈↓ α←␈↓solving the problem;



␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?Heuristic search:


␈↓"β␈↓ α←␈↓␈↓ β?␈↓↓object␈α∞level␈↓--knowledge␈α∞of␈α
the␈α∞search␈α∞space␈α
(e.g.,␈α∞for␈α∞game␈α∞playing,␈α
a
␈↓ α←␈↓legal move generator and other operations on the domain primitives),

␈↓"β␈↓ α←␈↓␈↓ β?␈↓↓meta␈α∀level␈↓--the␈α∀various␈α∀search␈α∀strategies␈α∀(alpha-beta,␈α∀branch␈α∪and
␈↓ α←␈↓bound, hill climbing, etc.).

␈↓ α←␈↓To␈α∩emphasize␈α∩the␈α∩generality␈α∩issue,␈α∩consider␈α∩the␈α∩concept␈α∩of␈α∩the␈α∩PKS␈α⊃set,
␈↓ α←␈↓defined␈α∩as␈α∩all␈α∪KSs␈α∩that␈α∩are␈α∩``plausibly␈α∪useful.'' ␈α∩What␈α∩in␈α∪fact␈α∩constitutes
␈↓ α←␈↓``plausibly␈α
useful''?␈α
 For␈α
the␈α
vast␈α
majority␈α
of␈α
systems,␈α
the␈α
answer␈α
is␈α
assumed␈α
to
␈↓ α←␈↓be␈αobvious--in␈αproduction␈αsystems␈αand␈α␈↓¬PLANNER␈↓-like␈αlanguages,␈αfor␈αinstance,␈αit
␈↓ α←␈↓is␈α⊂based␈α∂on␈α⊂pattern␈α∂matching␈α⊂and,␈α∂after␈α⊂all,␈α∂it's␈α⊂``obvious''␈α∂when␈α⊂a␈α∂pattern
␈↓ α←␈↓matches.␈α Such␈αknowledge␈αis␈α
thus␈αoften␈αburied␈αdeep␈α
within␈αa␈αsystem␈α(if␈α
it␈αis
␈↓ α←␈↓represented␈αexplicitly␈αat␈αall),␈αand␈αindeed␈αthe␈α␈↓¬PLANNER␈↓␈αpattern␈αmatcher␈αoperates
␈↓ α←␈↓at␈αa␈αvery␈αlow␈αlevel,␈αcompared␈αwith␈αother␈αfeatures␈αof␈αthe␈αlanguage.␈αWe␈αclaim,
␈↓ α←␈↓however,␈α
that␈αeven␈α
this␈αchunk␈α
of␈α
knowledge␈αshould␈α
be␈αregarded␈α
as␈αa␈α
strategy
␈↓ α←␈↓and should be made explicit in the system.
␈↓"β␈↓ α←␈↓␈↓ β?Among␈α∪the␈α∪advantages␈α∪would␈α∩be␈α∪flexibility.␈α∪ For␈α∪example,␈α∪in␈α∩an
␈↓ α←␈↓event-driven␈α⊂system␈α⊂using␈α⊂pattern-directed␈α⊂invocation␈α⊂of␈α⊂KSs␈α⊂(e.g.,␈α⊂␈↓¬PLANNER␈↓
␈↓ α←␈↓antecedent␈αtheorems),␈αonly␈αa␈αcertain␈αsubset␈αof␈αthe␈αKSs␈αare␈αconsidered--those
␈↓ α←␈↓with␈α
patterns␈α
containing␈α
obvious␈αdiscrepancies␈α
with␈α
current␈α
information␈αare
␈↓ α←␈↓ignored.␈α∩Consider,␈α∩however,␈α∩the␈α∪utility␈α∩of␈α∩checking␈α∩for␈α∪misinterpreted␈α∩or
␈↓ α←␈↓missing␈α∞data␈α∂in␈α∞a␈α∂noisy␈α∞domain␈α∞in␈α∂the␈α∞following␈α∂way: ␈α∞If␈α∂several␈α∞strategies
␈↓ α←␈↓indicated␈α
strongly␈αthat␈α
a␈αcertain␈α
KS␈α
would␈αbe␈α
useful,␈αbut␈α
its␈α
pattern␈αdoesn't
␈↓ α←␈↓currently␈α⊃match,␈α∩it␈α⊃may␈α∩be␈α⊃useful␈α∩to␈α⊃retrieve␈α∩that␈α⊃KS␈α∩nevertheless.␈α⊃ The
␈↓ α←␈↓discrepancies␈αbetween␈α
this␈αKS␈α
and␈αthe␈α
current␈αdata␈α
base␈αmight␈α
in␈αfact␈α
be␈αa
␈↓ α←␈↓useful␈α∞hint␈α
about␈α∞where␈α∞to␈α
check␈α∞for␈α∞missing␈α
or␈α∞misinterpreted␈α∞data.␈α
Thus,
␈↓ α←␈↓by␈α∂turning␈α∂this␈α∂``obvious''␈α∂chunk␈α∂of␈α∂knowledge--defining␈α∂what␈α∂it␈α∂means␈α∞to
␈↓ α←␈↓``match''--into␈α⊂an␈α⊂explicit␈α⊂strategy,␈α⊂we␈α∂obtain␈α⊂a␈α⊂highly␈α⊂flexible␈α⊂system,␈α∂one
␈↓ α←␈↓that␈α∞can␈α∞be␈α∞``tuned''␈α∞to␈α∞account␈α∞for␈α∞the␈α∞degree␈α∞of␈α∞noise␈α∞in␈α∞the␈α∞domain␈α
with
␈↓ α←␈↓relatively␈α∪little␈α∪trouble.␈α∪Fundamental␈α∪changes␈α∪in␈α∪system␈α∀performance␈α∪can
␈↓ α←␈↓therefore be controlled at a fairly high level.
␈↓ α←␈↓␈↓7-3␈↓ π{WHAT IS A STRATEGY    201␈↓

␈↓"β␈↓ α←␈↓␈↓ β?Note␈αthat␈α
this␈αview␈α
makes␈αfew␈α
assumptions␈αconcerning␈αthe␈α
underlying
␈↓ α←␈↓methodology.␈α↔ It␈α↔appears␈α⊗to␈α↔be␈α↔a␈α↔useful␈α⊗perspective␈α↔over␈α↔a␈α↔range␈α⊗of
␈↓ α←␈↓representation␈α→techniques.␈α→ It␈α→describes␈α→equally␈α→well,␈α→for␈α→example,␈α→the
␈↓ α←␈↓organization␈αof␈αknowledge␈αfor␈αsystems␈αthat␈αare␈αevent␈αdriven␈αas␈αwell␈αas␈αthose
␈↓ α←␈↓that␈α
are␈α
goal␈α
directed.␈α
 In␈α
the␈α
former␈α
case,␈α
the␈α
question␈α
is␈α
how␈α
to␈α
select␈α
the
␈↓ α←␈↓most␈α
relevant␈α
subset␈αfrom␈α
all␈α
the␈αpotential␈α
implications␈α
of␈αa␈α
new␈α
event;␈αin␈α
the
␈↓ α←␈↓latter␈αit␈αbecomes␈αthe␈αintelligent␈αchoice␈αof␈αa␈αKS␈αthat␈αhelps␈αachieve␈αthe␈αcurrent
␈↓ α←␈↓goal.

␈↓"β␈↓ α←␈↓␈↓αStrategies as a means of defining and choosing invocation criteria␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α⊂strategies␈α⊃are␈α⊂also␈α⊃traditionally␈α⊂viewed␈α⊂as␈α⊃a␈α⊂kind␈α⊃of␈α⊂``fine-
␈↓ α←␈↓tuning''␈α∪of␈α∪a␈α∩general␈α∪method;␈α∪as,␈α∪for␈α∩example,␈α∪set-of-support␈α∪is␈α∪used␈α∩to
␈↓ α←␈↓improve␈α⊃the␈α⊃performance␈α⊃of␈α⊂resolution␈α⊃theorem␈α⊃proving.␈α⊃ As␈α⊃will␈α⊂become
␈↓ α←␈↓clear␈αbelow,␈αhowever,␈αwe␈α
need␈αnot␈αrestrict␈αourselves␈α
to␈αthis␈αview.␈α In␈αthe␈α
most
␈↓ α←␈↓general␈α∂terms,␈α∂we␈α∂can␈α∂view␈α∞strategic␈α∂knowledge␈α∂as␈α∂any␈α∂decision␈α∞concerning
␈↓ α←␈↓how␈αor␈αwhich␈α
knowledge␈αto␈αuse.␈α
 Adopting␈αthis␈αview␈α
will␈αmake␈αit␈α
clear␈αthat
␈↓ α←␈↓strategies␈α
need␈α∞not␈α
be␈α∞set␈α
in␈α∞the␈α
context␈α
of␈α∞tuning␈α
a␈α∞single␈α
method␈α∞but␈α
can
␈↓ α←␈↓more␈α∂generally␈α∞be␈α∂used␈α∞to␈α∂choose␈α∞(or␈α∂even␈α∞define)␈α∂the␈α∞method␈α∂itself.␈α∞ This
␈↓ α←␈↓view is developed further in Section 7-5.

␈↓"β␈↓ α←␈↓␈↓α7-3-3    Levels of knowledge␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αdiscussion␈αthus␈αfar␈αhas␈αsuggested␈αthe␈αexistence␈αof␈αonly␈αtwo␈αlevels
␈↓ α←␈↓of␈α⊂knowledge,␈α⊂the␈α⊂object␈α⊂and␈α⊂meta␈α⊂levels.␈α⊂ It␈α⊂seems␈α⊂plausible,␈α⊂however,␈α⊂to
␈↓ α←␈↓continue␈α⊗this␈α⊗sequence␈α⊗through␈α⊗an␈α⊗arbitrary␈α⊗number␈α⊗of␈α↔levels.␈α⊗Where
␈↓ α←␈↓strategies␈α⊃direct␈α⊃the␈α⊂use␈α⊃of␈α⊃object-level␈α⊂knowledge,␈α⊃second␈α⊃order␈α⊂strategies
␈↓ α←␈↓direct␈αthe␈αuse␈αof␈αstrategies␈α(e.g.,␈αwhen␈αis␈αhill␈αclimbing␈αbetter␈αthan␈αbranch␈α
and
␈↓ α←␈↓bound),␈αthird␈αorder␈αstrategies␈αsuggest␈αthe␈αcriteria␈αfor␈αchoosing␈αstrategies␈α(e.g.,
␈↓ α←␈↓how␈α
do␈α
I␈αgo␈α
about␈α
deciding␈αwhich␈α
search␈α
technique␈α
is␈αthe␈α
best),␈α
and␈αso␈α
forth.
␈↓ α←␈↓Note␈α
that␈α
the␈αprocess␈α
is␈α
the␈α
same␈αat␈α
all␈α
levels;␈α
each␈αlevel␈α
directs␈α
the␈α
use␈αof␈α
the
␈↓ α←␈↓knowledge at the next lower level.␈↓
6␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Part␈αof␈αthe␈αintuitive␈αappeal␈αof␈αthe␈αhigher␈αlevels␈αof␈αstrategy␈αlies␈αin␈αthe
␈↓ α←␈↓belief␈α
that␈α
the␈α
invocation␈α
of␈α
a␈α∞useful␈α
piece␈α
of␈α
strategy␈α
knowledge␈α
at␈α∞a␈α
high
␈↓ α←␈↓level␈α∂offers␈α∞advantages␈α∂similar␈α∂to␈α∞choosing␈α∂an␈α∞appropriate␈α∂branch␈α∂early␈α∞in
␈↓ α←␈↓the␈α∞search␈α∞of␈α
a␈α∞large␈α∞tree.␈α
 That␈α∞is,␈α∞since␈α
the␈α∞strategy␈α∞at␈α
level␈α∞N␈α∞selects␈α
the
␈↓ α←␈↓relevant␈α⊃strategies␈α⊃at␈α⊃level␈α⊃N-1,␈α⊃and␈α⊃since␈α⊃this␈α⊃carefully␈α⊃chosen␈α⊃subset␈α⊃of
␈↓ α←␈↓strategies␈α∞does␈α∂the␈α∞same␈α∞for␈α∂the␈α∞strategies␈α∂at␈α∞level␈α∞N-2,␈α∂etc.,␈α∞the␈α∂strategy␈α∞at
␈↓ α←␈↓level N can exert a powerful focusing influence on the whole process.
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[6]␈α∂This␈α∂hierarchy␈α∂is␈α⊂analogous␈α∂to␈α∂the␈α∂one␈α⊂in␈α∂chapter␈α∂6␈α∂labeled␈α⊂``levels␈α∂of
␈↓ α←␈↓generality.''  In␈αboth␈αcases,␈αeach␈αhigher␈αlevel␈αdescribes␈αknowledge␈αat␈αthe␈αnext
␈↓ α←␈↓lower␈α⊂level.␈α⊂ Each␈α⊂level␈α⊂of␈α⊃the␈α⊂hierarchy␈α⊂discussed␈α⊂in␈α⊂chapter␈α⊃6␈α⊂(instance,
␈↓ α←␈↓schema,␈α∩schema-schema)␈α⊃describes␈α∩the␈α⊃␈↓↓representation␈↓␈α∩of␈α⊃knowledge␈α∩at␈α⊃the
␈↓ α←␈↓next␈α
lower␈α
level;␈αhere,␈α
each␈α
level␈αof␈α
strategy␈α
describes␈αthe␈α
␈↓↓use␈↓␈α
of␈αknowledge␈α
at
␈↓ α←␈↓the␈α∞next␈α
lower␈α∞level.␈α
However,␈α∞while␈α∞the␈α
hierarchy␈α∞of␈α
chapter␈α∞6␈α∞appears␈α
to
␈↓ α←␈↓end␈α⊂at␈α⊂the␈α⊃third␈α⊂level,␈α⊂here␈α⊃there␈α⊂appears␈α⊂to␈α⊂be␈α⊃no␈α⊂ a priori␈α⊂limit␈α⊃to␈α⊂the
␈↓ α←␈↓number of levels.
␈↓ α←␈↓␈↓202    STRATEGIES␈↓ 
#7-3␈↓

␈↓"β␈↓ α←␈↓␈↓ β?Two␈α
global␈α
considerations␈αare␈α
relevant␈α
here.␈α First,␈α
the␈α
domain␈αmust
␈↓ α←␈↓allow␈α
successive␈αlevels␈α
of␈αstrategies.␈α
 It␈α
is␈αnot␈α
obvious,␈αfor␈α
instance,␈αwhat␈α
some
␈↓ α←␈↓of␈α∞the␈α∞fourth␈α∞and␈α
higher␈α∞level␈α∞strategies␈α∞would␈α
look␈α∞like,␈α∞even␈α∞for␈α
familiar
␈↓ α←␈↓methodologies␈α
like␈α∞heuristic␈α
search.␈α
The␈α∞domain␈α
should␈α
thus␈α∞be␈α
sufficiently
␈↓ α←␈↓formalized␈α
that␈α
it␈α
is␈α
possible␈αto␈α
determine␈α
the␈α
conceptual␈α
primitives␈α
of␈αeach
␈↓ α←␈↓level of knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?Second,␈α⊗these␈α∃strategies␈α⊗must␈α⊗be␈α∃effective,␈α⊗that␈α∃is,␈α⊗the␈α⊗result␈α∃of
␈↓ α←␈↓invoking␈αany␈αone␈αlevel␈αof␈αstrategy␈αmust␈αbe␈αa␈αuseful␈αselection␈αof␈αknowledge␈αat
␈↓ α←␈↓the␈α
next␈α
lower␈α
level.␈α
This␈α
means␈α
that␈α
some␈α
fairly␈α
specific␈α
statements␈α
can␈αbe
␈↓ α←␈↓made about useful choices at each level.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α
what␈α
follows,␈αfor␈α
the␈α
sake␈α
of␈αexposition,␈α
most␈α
issues␈α
are␈αdiscussed
␈↓ α←␈↓as␈α
if␈αthere␈α
were␈αonly␈α
two␈α
levels,␈αbut␈α
they␈αcan␈α
be␈αgeneralized␈α
to␈α
an␈αarbitrary
␈↓ α←␈↓number of levels.
␈↓"β␈↓ α←␈↓␈↓ β?By␈α⊃combining␈α⊂the␈α⊃concept␈α⊂of␈α⊃multiple␈α⊂levels␈α⊃with␈α⊂the␈α⊃view␈α⊃of␈α⊂the
␈↓ α←␈↓nature␈α∂of␈α⊂strategic␈α∂knowledge␈α∂suggested␈α⊂earlier,␈α∂we␈α∂see␈α⊂the␈α∂beginning␈α⊂of␈α∂a
␈↓ α←␈↓general␈α≤framework␈α≤for␈α≤the␈α≤organization␈α≤and␈α≤explication␈α≤of␈α≠strategy
␈↓ α←␈↓knowledge.␈α
 The␈α
selection␈α
of␈α
a␈αmethodology␈α
(like␈α
the␈α
ones␈α
in␈α
Section␈α7-3-2)
␈↓ α←␈↓might␈α
for␈αinstance␈α
be␈αaccomplished␈α
by␈αa␈α
hierarchical␈αset␈α
of␈α
strategies␈αwhich
␈↓ α←␈↓decide,␈α∂first,␈α∂␈↓↓how␈α∂to␈α∂choose␈↓␈α∂a␈α∂methodology␈α∂(second␈α∂order␈α∂strategy)␈α∂and␈α∞then
␈↓ α←␈↓␈↓↓which␈α∀methodology␈↓␈α∀to␈α∀choose␈α∀(first-order).␈α∀ As␈α∀before,␈α∀the␈α∀selection␈α∀of␈α∪a
␈↓ α←␈↓particular␈α⊂methodology␈α⊂may␈α∂then␈α⊂imply␈α⊂further␈α∂decisions␈α⊂to␈α⊂be␈α⊂made␈α∂that
␈↓ α←␈↓would once again be the task of strategic knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?While␈α∂writing␈α∂strategies␈α∂can␈α∂involve␈α∂a␈α∂significant␈α∂amount␈α∂of␈α∂effort,
␈↓ α←␈↓even␈αa␈α
moderate␈αinvestment␈α
appears␈αto␈αpay␈α
off␈αquite␈α
well.␈α Certain␈αaspects␈α
of
␈↓ α←␈↓the␈αorganization␈αof␈αthe␈α␈↓¬HEARSAY␈αII␈↓␈αsystem,␈αfor␈αexample,␈αcan␈αbe␈αviewed␈αin␈αlight
␈↓ α←␈↓of␈αsuccessive␈αlevels␈αof␈αknowledge,␈αdemonstrating␈αthe␈αpower␈αof␈α
the␈αtechnique.
␈↓ α←␈↓That␈αsystem␈αis␈αbuilt␈α
around␈αa␈αlarge␈αnumber␈α
of␈αdemons,␈αeach␈αwaiting␈α
to␈αfire
␈↓ α←␈↓and␈α
contribute␈α
its␈αchunk␈α
of␈α
knowledge.␈αEach␈α
demon␈α
has␈αa␈α
set␈α
of␈αarbitrarily
␈↓ α←␈↓complex␈αpreconditions␈αwhich␈αspecify␈α
the␈αcircumstances␈αnecessary␈αfor␈αit␈α
to␈αbe
␈↓ α←␈↓relevant.␈α
 But␈αit␈α
would␈αbe␈α
prohibitively␈αexpensive␈α
to␈αevaluate␈α
all␈αimplicated
␈↓ α←␈↓preconditions␈α⊃every␈α⊂time␈α⊃a␈α⊂new␈α⊃datum␈α⊃enters␈α⊂the␈α⊃data␈α⊂base.␈α⊃ In␈α⊃order␈α⊂to
␈↓ α←␈↓avoid␈α⊂this,␈α⊂the␈α∂precondition␈α⊂of␈α⊂each␈α⊂demon␈α∂has␈α⊂its␈α⊂own␈α⊂precondition␈α∂(the
␈↓ α←␈↓``pre-precondition''),␈α∞which␈α∞specifies␈α∞the␈α∞general␈α∞conditions␈α∞under␈α∞which␈α
the
␈↓ α←␈↓system␈α∞should␈α∞bother␈α∞to␈α∞evaluate␈α∞the␈α∞whole␈α∞precondition.␈α∞ Apparently␈α∞even
␈↓ α←␈↓these␈αtwo␈αlevels␈αof␈αknowledge␈αare␈αsufficiently␈αpowerful␈αto␈αsupport␈αacceptable
␈↓ α←␈↓performance.

␈↓"β␈↓ α←␈↓␈↓α7-3-4    Building blocks:  Conceptual primitives, language␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Object-level␈αknowledge␈αis␈αbuilt␈αfrom␈αconceptual␈αprimitives␈αspecific␈αto
␈↓ α←␈↓the␈α∂domain␈α∂of␈α⊂application.␈α∂ In␈α∂␈↓¬MYCIN␈↓,␈α∂for␈α⊂instance,␈α∂the␈α∂notions␈α⊂of␈α∂organism
␈↓ α←␈↓gramstain,␈α⊗morphology,␈α⊗and␈α⊗identity␈α⊗are␈α⊗relevant␈α⊗concepts.␈α∃Object-level
␈↓ α←␈↓primitives␈α↔thus␈α⊗characterize␈α↔the␈α↔domain,␈α⊗and␈α↔it␈α↔is␈α⊗the␈α↔search␈α↔for␈α⊗an
␈↓ α←␈↓appropriate␈α
set␈α
of␈α
them,␈α
and␈α
a␈αlanguage␈α
in␈α
which␈α
to␈α
express␈α
and␈α
use␈αthem,
␈↓ α←␈↓that is a large part of the traditional representation problem of AI.
␈↓"β␈↓ α←␈↓␈↓ β?Strategies,␈α∀on␈α∀the␈α∃other␈α∀hand,␈α∀require␈α∀conceptual␈α∃primitives␈α∀that
␈↓ α←␈↓␈↓7-3␈↓ π{WHAT IS A STRATEGY    203␈↓

␈↓"β␈↓ α←␈↓describe␈α↔characteristics␈α⊗of␈α↔knowledge␈α↔rather␈α⊗than␈α↔characteristics␈α↔of␈α⊗the
␈↓ α←␈↓domain.␈α
 Some␈α
of␈α
those␈α
primitives␈αmight␈α
deal␈α
with␈α
general␈α
attributes␈α
of␈αa␈α
KS
␈↓ α←␈↓like:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?preconditions for its use--to assure its utility,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?any side effects--to make clear all the implications of using it,
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?its main effect--so that it can be used when relevant, and
␈↓"β␈↓ α←␈↓␈↓ ββ(d)␈↓ β?the␈α∪main␈α∪factors␈α∪on␈α∪which␈α∪it␈α∪is␈α∪based--a␈α∪finer␈α∀degree␈α∪of
␈↓ α←␈↓␈↓ β?characterization than main effect.

␈↓"β␈↓ α←␈↓␈↓ β?Others␈α⊗might␈α⊗be␈α⊗suggested␈α⊗by␈α⊗the␈α⊗structure␈α⊗of␈α⊗the␈α⊗object-level
␈↓ α←␈↓primitives.␈α
 For␈α∞instance,␈α
as␈α
noted␈α∞earlier,␈α
the␈α
current␈α∞performance␈α
program
␈↓ α←␈↓uses␈α∪both␈α∪consequent␈α∀and␈α∪antecedent␈α∪rules;␈α∪each␈α∀rule␈α∪is␈α∪composed␈α∀of␈α∪a
␈↓ α←␈↓premise␈αand␈αan␈α
action;␈αthese,␈αin␈αturn,␈α
are␈αmade␈αup␈α
of␈αclauses;␈αand␈αthe␈α
clauses
␈↓ α←␈↓are␈αbuilt␈αfrom␈α
predicate␈αfunctions,␈αattributes,␈α
objects,␈αvalues,␈αcertainty␈α
factors,
␈↓ α←␈↓etc.␈α Each␈αof␈αthese␈αsuggests␈α
several␈αpossible␈αmeta-level␈αprimitives: ␈αIs␈αthe␈α
rule
␈↓ α←␈↓an␈α∩antecedent␈α∩or␈α∩consequent␈α∩rule?␈α∩How␈α∩many␈α∩premise␈α∩clauses␈α∩are␈α⊃there?
␈↓ α←␈↓Which␈αfunctions,␈αattributes,␈αetc.,␈αdoes␈α
the␈αrule␈αemploy?␈α Is␈αthe␈αcertainty␈α
factor
␈↓ α←␈↓positive␈αor␈α
negative?␈αetc.␈α
 In␈αthe␈αsame␈α
way,␈αthe␈α
components␈αof␈α
any␈αsort␈αof␈α
KS
␈↓ α←␈↓could provide hints about potentially useful primitives for characterizing it.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∂suitably␈α∂large␈α∂set␈α∂of␈α⊂such␈α∂primitives␈α∂would␈α∂form␈α∂the␈α∂basis␈α⊂for␈α∂a
␈↓ α←␈↓useful␈α␈↓↓strategy␈αlanguage␈↓.␈αMore␈α
is␈αof␈αcourse␈αneeded: ␈α
It␈αis␈αnot␈αenough␈α
just␈αto
␈↓ α←␈↓choose␈α∂the␈α⊂primitives;␈α∂there␈α⊂must␈α∂be␈α⊂a␈α∂way␈α⊂of␈α∂expressing,␈α⊂combining,␈α∂and
␈↓ α←␈↓using␈α∩them,␈α⊃so␈α∩as␈α⊃to␈α∩effectively␈α⊃direct␈α∩the␈α⊃use␈α∩of␈α∩object-level␈α⊃knowledge.
␈↓ α←␈↓Since␈α⊂there␈α⊂are␈α⊂many␈α∂advantages␈α⊂to␈α⊂be␈α⊂gained␈α∂by␈α⊂a␈α⊂uniform␈α⊂encoding␈α∂of
␈↓ α←␈↓knowledge,␈α∂it␈α∂is␈α∞useful␈α∂if␈α∂the␈α∂object-level␈α∞syntax␈α∂can␈α∂simply␈α∂be␈α∞augmented
␈↓ α←␈↓with␈α∩the␈α∩new␈α∩meta-level␈α∩primitives␈α∩to␈α∩provide␈α∩the␈α∩strategy␈α∩language.␈α⊃ A
␈↓ α←␈↓demonstration␈α
of␈α
this␈α
technique␈α
and␈α
a␈α
discussion␈α
of␈α
its␈α
advantages␈α∞is␈α
given
␈↓ α←␈↓below.
␈↓"β␈↓ α←␈↓␈↓ β?While␈αthis␈αview␈αhas␈αindicated␈αwhere␈αto␈αfind␈αa␈αuseful␈αsubset␈αof␈αmeta-
␈↓ α←␈↓level␈α∞primitives,␈α∞the␈α∞question␈α∞of␈α∞determining␈α∞the␈α∞entire␈α∞collection␈α∞may␈α∂be␈α∞a
␈↓ α←␈↓good␈α∂deal␈α∂harder.␈α∞ We␈α∂claim␈α∂below␈α∂that␈α∞the␈α∂set␈α∂is␈α∞at␈α∂least␈α∂very␈α∂large,␈α∞and
␈↓ α←␈↓perhaps␈αopen-ended.␈α These␈αclaims␈αare,␈αin␈αturn,␈αimportant␈αconsiderations␈αfor
␈↓ α←␈↓the design of a strategy representation and are discussed in Section 7-5-3.
␈↓ α←␈↓␈↓204    STRATEGIES␈↓ 
#7-4␈↓

␈↓"β␈↓ α←␈↓␈↓α7-4    META-RULES␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈α∪noted,␈α∩the␈α∪initial␈α∩motivation␈α∪for␈α∩meta-rules␈α∪was␈α∩to␈α∪provide␈α∩a
␈↓ α←␈↓mechanism␈α∞to␈α∞guide␈α∞the␈α∂performance␈α∞program␈α∞faced␈α∞with␈α∂saturation.␈α∞ This
␈↓ α←␈↓section␈α∞discusses␈α∞issues␈α∞of␈α∂meta-rule␈α∞design␈α∞and␈α∞representation␈α∂primarily␈α∞in
␈↓ α←␈↓that␈α∃light.␈α∃ It␈α∃also␈α∃considers␈α∃additional␈α∃applications␈α∃for␈α∃meta-rules␈α∀and
␈↓ α←␈↓demonstrates␈α∪their␈α∪utility␈α∪in␈α∪a␈α∪variety␈α∪of␈α∪contexts.␈α∪ Shortcomings␈α∀of␈α∪the
␈↓ α←␈↓current implementation are also reviewed.

␈↓"β␈↓ α←␈↓␈↓α7-4-1    Format␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Two␈α∂examples␈α∞of␈α∂meta-rules␈α∂are␈α∞shown␈α∂below,␈α∞in␈α∂both␈α∂the␈α∞internal
␈↓ α←␈↓format and the English translation that results.␈↓
7␈↓

␈↓"β␈↓ α←␈↓	␈↓&METARULE001␈↓)αβ

␈↓"β␈↓ α←␈↓	If   1) the culture was not obtained from a sterile source, and
␈↓"β␈↓ α←␈↓	     2) there are rules which mention in their premise a
␈↓"β␈↓ α←␈↓	        previous organism which may be the same as the current
␈↓"β␈↓ α←␈↓	        organism
␈↓"β␈↓ α←␈↓	Then it is definite (1.0) that each of them is not going to be
␈↓"β␈↓ α←␈↓	useful.

␈↓"β␈↓ α←␈↓	PREMISE: ($AND(NOTSAME CNTXT STERILESOURCE)
␈↓"β␈↓ α←␈↓	              (THEREARE OBJRULES
␈↓"β␈↓ α←␈↓	                       (MENTIONS CNTXT PREMISE SAMEBUG) SET1))
␈↓"β␈↓ α←␈↓	ACTION:  (CONCLUDE SET1 UTILITY NO TALLY 1.0)

␈↓"β␈↓ α←␈↓	␈↓&METARULE002␈↓)αβ

␈↓"β␈↓ α←␈↓	If   1) the infection is a pelvic-abscess, and
␈↓"β␈↓ α←␈↓	     2) there are rules which mention in their conclusion
␈↓"β␈↓ α←␈↓	        enterobacteriaceae, and
␈↓"β␈↓ α←␈↓	     3) there are rules which mention in their conclusion
␈↓"β␈↓ α←␈↓	        grampos-rods,
␈↓"β␈↓ α←␈↓	There is suggestive evidence (.4) that the former should be
␈↓"β␈↓ α←␈↓	done before the latter.

␈↓"β␈↓ α←␈↓	PREMISE: ($AND (SAME CNTXT PELVIC-ABSCESS)
␈↓"β␈↓ α←␈↓	               (THEREARE OBJRULES(MENTIONS CNTXT ACTION
␈↓"β␈↓ α←␈↓	                                  ENTEROBACTERIACEAE) SET1)
␈↓"β␈↓ α←␈↓	               (THEREARE OBJRULES(MENTIONS CNTXT ACTION
␈↓"β␈↓ α←␈↓	                                  GRAMPOS-RODS) SET2))
␈↓"β␈↓ α←␈↓	ACTION:  (CONCLUDE SET1 DOBEFORE SET2 TALLY .4)


␈↓"β␈↓ α←␈↓α␈↓ ∧tFig. 7-1.    Meta-rule examples.    

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[7]␈α⊂For␈α⊂reasons␈α⊂noted␈α⊂in␈α∂chapter␈α⊂1,␈α⊂the␈α⊂translation␈α⊂is␈α⊂somewhat␈α∂awkward;
␈↓ α←␈↓these␈αparticular␈αrules␈αsuffer␈αespecially␈αfrom␈αthe␈αclause-by-clause␈αapproach␈αto
␈↓ α←␈↓translation.
␈↓"β␈↓ α←␈↓␈↓7-4␈↓ λbMETA-RULES    205␈↓

␈↓"β␈↓ α←␈↓The␈α
strategic␈α
impact␈α
of␈α
the␈α
first␈α
rule␈α
arises␈α
from␈α
the␈α
fact␈α
that␈α
an␈α
old␈α
infection
␈↓ α←␈↓which␈α⊂has␈α∂been␈α⊂cured␈α∂only␈α⊂temporarily␈α⊂may␈α∂recur,␈α⊂perhaps␈α∂as␈α⊂much␈α⊂as␈α∂a
␈↓ α←␈↓month␈α⊂later.␈α⊃ Thus,␈α⊂one␈α⊂of␈α⊃the␈α⊂possible␈α⊃ways␈α⊂to␈α⊂deduce␈α⊃the␈α⊂identity␈α⊃of␈α⊂a
␈↓ α←␈↓current␈α
organism␈αis␈α
by␈αreference␈α
to␈αprevious␈α
infections.␈α However,␈α
this␈αline␈α
of
␈↓ α←␈↓reasoning␈αis␈αnot␈αvalid␈αif␈αthe␈αcurrent␈αinfection␈αwas␈αcultured␈αin␈αa␈αfashion␈αthat
␈↓ α←␈↓caused␈α
the␈α
sample␈α
to␈α
be␈α
nonsterile␈α
(e.g.,␈α
taken␈α
from␈α
a␈α
nonsterile␈α
site).␈α Thus
␈↓ α←␈↓the␈αrule,␈αin␈αeffect,␈αsays␈α␈↓↓if␈αthe␈αcurrent␈αculture␈αis␈αnot␈αfrom␈αa␈αsterile␈αsource,␈αdon't
␈↓ α←␈↓↓bother␈α∩trying␈α∩to␈α⊃deduce␈α∩the␈α∩current␈α∩organism␈α⊃identity␈α∩from␈α∩the␈α∩identity␈α⊃of
␈↓ α←␈↓↓previous␈αorganisms␈↓.␈α As␈αshould␈αbe␈αclear,␈αthis␈αis␈αa␈αglobal␈αstatement␈αabout␈αhow
␈↓ α←␈↓to␈α
reason␈α
in␈α
a␈α∞given␈α
situation␈α
and␈α
as␈α
such␈α∞offers␈α
a␈α
useful␈α
piece␈α∞of␈α
strategic
␈↓ α←␈↓advice.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αsecond␈αrule␈αindicates␈αthat␈αsince␈αenterobacteriaceae␈αare␈αcommonly
␈↓ α←␈↓associated␈α⊂with␈α⊂a␈α⊂pelvic␈α⊂abscess,␈α⊂it␈α⊂is␈α⊂a␈α⊂good␈α⊂idea␈α⊂to␈α⊂try␈α⊃rules␈α⊂concluding
␈↓ α←␈↓about them first, before the less likely grampositive rods.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αsyntax␈αfor␈αmeta-rules␈αis␈αidentical␈αto␈αobject-rule␈αsyntax,␈αextended
␈↓ α←␈↓to␈α∃include␈α∃two␈α∃new␈α∀predicate␈α∃functions␈α∃(␈↓	MENTIONS␈↓␈α∃and␈α∃its␈α∀complement,
␈↓ α←␈↓␈↓	DOESNTMENTION␈↓),␈α⊂and␈α∂two␈α⊂new␈α∂attributes␈α⊂(which␈α∂concern␈α⊂a␈α⊂rule's␈α∂␈↓	UTILITY␈↓
␈↓ α←␈↓and␈α∩its␈α⊃place␈α∩in␈α⊃the␈α∩sequence␈α∩of␈α⊃rules␈α∩to␈α⊃be␈α∩invoked␈α∩(␈↓	DOBEFORE␈↓)).␈α⊃ Two
␈↓ α←␈↓important␈α
benefits␈α
accrue␈α
from␈α
using␈αa␈α
single␈α
syntax: ␈α
First,␈α
meta-rules␈αmay
␈↓ α←␈↓employ␈α∃all␈α∃the␈α∀existing␈α∃machinery␈α∃of␈α∀certainty␈α∃factors␈α∃to␈α∃make␈α∀inexact
␈↓ α←␈↓statements,␈α∞and,␈α∞second,␈α∂the␈α∞system␈α∞has␈α∞a␈α∂uniform␈α∞encoding␈α∞of␈α∞all␈α∂levels␈α∞of
␈↓ α←␈↓knowledge.  Implications of these are discussed below.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂current␈α∂implementation␈α⊂has␈α⊂a␈α∂very␈α⊂simple␈α∂syntax␈α⊂but␈α⊂offers␈α∂a
␈↓ α←␈↓useful range of expression.  Meta-rules can indicate:

␈↓"β␈↓ α←␈↓	␈↓(a) the utility of object-level rules (as in METARULE001):␈↓	

␈↓"β␈↓ α←␈↓	        under conditions A and B,

␈↓"β␈↓ α←␈↓	        rules which do {not} mention X in their {premise
␈↓"β␈↓ α←␈↓	                                                 action}
␈↓"β␈↓ α←␈↓	        will {definitely be useless
␈↓"β␈↓ α←␈↓	              probably be useless
␈↓"β␈↓ α←␈↓	                ...
␈↓"β␈↓ α←␈↓	              probably be especially useful
␈↓"β␈↓ α←␈↓	              definitely be especially useful}
␈↓ α←␈↓␈↓206    STRATEGIES␈↓ 
#7-4␈↓

␈↓"β␈↓ α←␈↓	␈↓(b) a partial ordering of the object-level rules (as in METARULE002):␈↓	

␈↓"β␈↓ α←␈↓	        under conditions A and B,

␈↓"β␈↓ α←␈↓	        rules which do {not} mention X in their {premise
␈↓"β␈↓ α←␈↓	                                                 action}
␈↓"β␈↓ α←␈↓	        should {definitely      be used {first.
␈↓"β␈↓ α←␈↓	                probably                 last.
␈↓"β␈↓ α←␈↓	                 ...                     before
␈↓"β␈↓ α←␈↓	                possibly}                after}

␈↓"β␈↓ α←␈↓	        rules which do {not} mention Y in their {premise
␈↓"β␈↓ α←␈↓	                                                 action}

␈↓"β␈↓ α←␈↓␈↓ β?Note␈αthat␈αthe␈αprimitives␈αthat␈αhave␈αbeen␈αadded␈αto␈αthe␈αsyntax␈α
deal,␈αas
␈↓ α←␈↓expected,␈αwith␈α
characteristics␈αof␈α
knowledge.␈α For␈α
example,␈αthe␈α
new␈αpredicate
␈↓ α←␈↓functions␈α∀are␈α∪based␈α∀on␈α∪the␈α∀fact␈α∪that␈α∀the␈α∪knowledge␈α∀representation␈α∀is␈α∪a
␈↓ α←␈↓composite␈αstructure␈αwhose␈αcomponents␈αare␈αaccessible;␈αhence,␈αit␈αmakes␈αsense␈αto
␈↓ α←␈↓refer␈α∃to␈α∃the␈α∃content␈α∃of␈α∃a␈α∃rule.␈α∃ The␈α∃new␈α∃attributes␈α∃are␈α∃based␈α∃on␈α∃the
␈↓ α←␈↓recognition,␈αfirst,␈αthat␈αa␈αrule␈αinvocation␈αis␈αa␈αdiscrete␈αevent,␈αso␈αit␈αmakes␈αsense
␈↓ α←␈↓to␈α∪indicate␈α∀order;␈α∪and,␈α∀second,␈α∪that␈α∪one␈α∀rule␈α∪may␈α∀be␈α∪more␈α∀useful␈α∪than
␈↓ α←␈↓another,␈α∞so␈α∞we␈α∞can␈α∞talk␈α∞meaningfully␈α∞about␈α∞utility.␈α∞ These␈α∞simple␈α∞additions
␈↓ α←␈↓allow␈α∂the␈α∞statement␈α∂of␈α∞the␈α∂fairly␈α∞broad␈α∂range␈α∞of␈α∂strategies␈α∂indicated␈α∞above
␈↓ α←␈↓and have so far met all current needs.

␈↓"β␈↓ α←␈↓␈↓α7-4-2    Function␈↓
␈↓"β␈↓ α←␈↓␈↓ β?We␈α∃present␈α∀first␈α∃a␈α∀simplified␈α∃picture␈α∀of␈α∃meta-rule␈α∃function␈α∀and
␈↓ α←␈↓elaborate␈αon␈α
it␈αin␈α
several␈αstages.␈α In␈α
the␈αsimplest␈α
case,␈αmeta-rules␈α(like␈α
object-
␈↓ α←␈↓level␈α∩rules)␈α∪are␈α∩associated␈α∪with␈α∩an␈α∪attribute.␈α∩As␈α∪explained␈α∩in␈α∪chapter␈α∩2,
␈↓ α←␈↓during␈α∞the␈α∞course␈α∞of␈α
attempting␈α∞to␈α∞establish␈α∞the␈α
value␈α∞of␈α∞any␈α∞attribute,␈α
the
␈↓ α←␈↓system␈α⊂retrieves␈α⊂the␈α⊂list␈α⊂of␈α⊂rules␈α∂relevant␈α⊂to␈α⊂that␈α⊂attribute.␈α⊂ Before␈α⊂any␈α∂of
␈↓ α←␈↓these␈α
rules␈α
is␈α∞invoked,␈α
however,␈α
the␈α∞system␈α
checks␈α
for␈α∞meta-rules␈α
associated
␈↓ α←␈↓with␈αthe␈αsame␈αattribute.␈α If␈αthere␈αare␈αany,␈αthese␈αare␈αexecuted␈αfirst␈αand␈αact␈αto
␈↓ α←␈↓reorder␈α∞or␈α∞prune␈α∞the␈α∞list␈α∞of␈α∞object-level␈α∞rules␈α∞(thereby␈α∞guiding␈α∞the␈α∞system's
␈↓ α←␈↓search␈α∪through␈α∪the␈α∪goal␈α∪tree).␈α∪ The␈α∪modified␈α∪list␈α∪is␈α∪then␈α∪passed␈α∪to␈α∩the
␈↓ α←␈↓standard rule interpreter described in chapter 2.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∪is␈α∩no␈α∪reason␈α∩to␈α∪constrain␈α∩this␈α∪process␈α∩to␈α∪a␈α∩single␈α∪level␈α∩of
␈↓ α←␈↓strategies,␈αand␈αthe␈αcurrent␈αimplementation␈α
is␈αgeneral␈αis␈αthis␈αrespect.␈α
 If␈αfirst-
␈↓ α←␈↓order␈α
meta-rules␈α
are␈α∞present,␈α
the␈α
process␈α∞recurs,␈α
and␈α
second-order␈α∞rules␈α
are
␈↓ α←␈↓sought,␈α∞which␈α∞would␈α∞then␈α∞be␈α∞used␈α∞to␈α∞reorder␈α∞or␈α∞select␈α∞from␈α∞the␈α∞first-order
␈↓ α←␈↓list,␈αand␈αso␈αon.␈↓
8␈↓␈αRecursion␈αstops␈αwhen␈αthere␈αis␈αno␈αrule␈αset␈αof␈αthe␈αnext␈αhigher

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[8]␈α
We␈α
have␈α
not␈αas␈α
yet␈α
uncovered␈α
any␈αsecond␈α
or␈α
higher␈α
order␈αmeta-rules,␈α
but
␈↓ α←␈↓then␈α
neither␈αhave␈α
we␈αactively␈α
looked␈αfor␈α
them.␈α In␈α
general,␈α
meta-rules␈αhave
␈↓ α←␈↓offered more expressive power than we have yet been able to use.
␈↓ α←␈↓␈↓7-4␈↓ λbMETA-RULES    207␈↓

␈↓"β␈↓ α←␈↓order,␈α∞and␈α∞the␈α∞process␈α∞unwinds,␈α∂allowing␈α∞each␈α∞level␈α∞of␈α∞knowledge␈α∂to␈α∞direct
␈↓ α←␈↓the use of the next lower level.
␈↓"β␈↓ α←␈↓␈↓ β?Meta-rules␈α∂may␈α∞also␈α∂be␈α∞written␈α∂to␈α∞control␈α∂the␈α∞invocation␈α∂of␈α∞object-
␈↓ α←␈↓level␈α
antecedent␈α
rules.␈α
 When␈α
a␈α
conclusion␈α
is␈α
made,␈α
the␈α
system␈α∞retrieves␈α
the
␈↓ α←␈↓list␈α⊃of␈α⊃antecedent␈α⊃rules␈α⊃associated␈α⊃with␈α⊃that␈α⊃conclusion.␈α⊃ If␈α⊃the␈α⊃list␈α∩is␈α⊃not
␈↓ α←␈↓empty,␈α
the␈α
system␈α
checks␈α
for␈αthe␈α
existence␈α
of␈α
applicable␈α
meta-rules␈α(i.e.,␈α
meta-
␈↓ α←␈↓rules␈α
associated␈α∞with␈α
the␈α∞antecedent␈α
rules␈α
dealing␈α∞with␈α
that␈α∞conclusion)␈α
and
␈↓ α←␈↓allows␈α⊂them␈α⊂to␈α⊂reorder␈α⊂or␈α⊂prune␈α⊂the␈α⊂list␈α⊂of␈α⊂conclusions␈α⊂to␈α⊂be␈α⊂made.␈α⊂This
␈↓ α←␈↓provides␈αa␈α
mechanism␈αfor␈α
writing␈αstrategies␈α
to␈αcontrol␈α
the␈αdepth␈αand␈α
breadth
␈↓ α←␈↓of implications drawn from any new fact or conclusion.

␈↓"β␈↓ α←␈↓␈↓αDetails␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Meta-rules␈α∃operate␈α∃by␈α∃making␈α∃conclusions␈α∃about␈α∃the␈α⊗utility␈α∃and
␈↓ α←␈↓relative␈αordering␈αfor␈αeach␈αobject-level␈αrule.␈α
To␈αsee␈αhow␈αthis␈αis␈αdone,␈α
consider
␈↓ α←␈↓the invocation of the two meta-rules shown earlier in Fig. 7-1.
␈↓"β␈↓ α←␈↓␈↓ β?Assume␈α∪the␈α∪system␈α∩is␈α∪attempting␈α∪to␈α∩determine␈α∪the␈α∪identity␈α∪of␈α∩an
␈↓ α←␈↓organism.␈α∞ It␈α∞will␈α∞retrieve␈α∞the␈α∞list␈α∞of␈α∞all␈α∞object-level␈α∞rules␈α∞concluding␈α
about
␈↓ α←␈↓identity␈α∪(call␈α∪the␈α∩list␈α∪L)␈α∪and␈α∩then␈α∪the␈α∪list␈α∩of␈α∪meta-rules␈α∪associated␈α∩with
␈↓ α←␈↓identity.␈α⊂Assume␈α⊂the␈α∂process␈α⊂ends␈α⊂here␈α∂because␈α⊂there␈α⊂are␈α⊂no␈α∂second-order
␈↓ α←␈↓meta-rules␈α⊂and␈α⊂that␈α∂the␈α⊂first␈α⊂meta-rule␈α∂in␈α⊂the␈α⊂list␈α∂is␈α⊂␈↓	METARULE001␈↓.␈α⊂If␈α∂the
␈↓ α←␈↓culture␈α∪under␈α∪consideration␈α∩is␈α∪not␈α∪from␈α∪a␈α∩sterile␈α∪source,␈α∪the␈α∪first␈α∩clause
␈↓ α←␈↓succeeds. Evaluation of the second clause

␈↓"β␈↓ α←␈↓	␈↓ β∂(THEREARE OBJRULES(MENTIONS CNTXT PREMISE SAMEBUG) SET1))

␈↓ α←␈↓results␈α⊃in␈α⊂assigning␈α⊃to␈α⊂␈↓	SET1␈↓␈α⊃the␈α⊂subset␈α⊃of␈α⊂rules␈α⊃that␈α⊂mention␈α⊃``a␈α⊂previous
␈↓ α←␈↓organism with possibly the same identity as the current organism'' (␈↓	SAMEBUG␈↓).
␈↓"β␈↓ α←␈↓␈↓ β?Note␈α
that␈α
determining␈α
the␈α∞appropriate␈α
subset␈α
of␈α
object-level␈α∞rules␈α
is
␈↓ α←␈↓accomplished␈α⊃by␈α⊃direct␈α⊃examination␈α⊃of␈α⊂the␈α⊃code␈α⊃of␈α⊃the␈α⊃object-level␈α⊂rules.
␈↓ α←␈↓That␈α∂is,␈α∞each␈α∂of␈α∞the␈α∂rules␈α∂in␈α∞L␈α∂is␈α∞tested␈α∂by␈α∞the␈α∂function␈α∂␈↓	MENTIONS␈↓,␈α∞which
␈↓ α←␈↓examines␈α
the␈αsource␈α
code␈αof␈α
the␈αrule␈α
to␈αsee␈α
(in␈αthis␈α
case)␈αif␈α
that␈αrule␈α
mentions
␈↓ α←␈↓the␈α≤attribute␈α≤␈↓	SAMEBUG␈↓␈α≤in␈α≤its␈α≤␈↓	PREMISE␈↓.␈α≤ (Implications␈α≤of␈α≤this␈α≤direct
␈↓ α←␈↓examination␈α∂of␈α∂source␈α⊂code--which␈α∂we␈α∂call␈α⊂␈↓↓content-directed␈α∂invocation␈↓--are
␈↓ α←␈↓explored below in Section 7-5.)
␈↓"β␈↓ α←␈↓␈↓ β?The␈α_action␈α_part␈α_of␈α→␈↓	METARULE001␈↓␈α_then␈α_concludes␈α_that␈α→each␈α_is
␈↓ α←␈↓definitely not useful.
␈↓"β␈↓ α←␈↓␈↓ β?Evaluating␈α∩␈↓	METARULE002␈↓␈α∩proceeds␈α∩analogously:␈α⊃If␈α∩the␈α∩patient␈α∩is␈α⊃a
␈↓ α←␈↓compromised␈α∪host,␈α∪␈↓	SET1␈↓␈α∪is␈α∪assigned␈α∪the␈α∪list␈α∪of␈α∪all␈α∪rules␈α∪mentioning␈α∩the
␈↓ α←␈↓identity␈α∞pseudomonas,␈α∞␈↓	SET2␈↓␈α∞is␈α∞assigned␈α∞all␈α∞rules␈α∞mentioning␈α∞klebsiella.␈α∞The
␈↓ α←␈↓conclusion␈α
would␈α
then␈α
indicate␈α
that␈α
there␈α
is␈α
``suggestive␈α
evidence''␈α∞that␈α
each
␈↓ α←␈↓rule␈α
in␈α
␈↓	SET1␈↓␈α
should␈α
be␈α
done␈α
before␈α
any␈α
of␈α
those␈α
in␈α
␈↓	SET2␈↓.␈α
This␈α
employs␈αall
␈↓ α←␈↓the␈α⊂pre-existing␈α⊂certainty␈α⊂factor␈α∂machinery,␈α⊂allowing␈α⊂the␈α⊂curious␈α⊂ability␈α∂to
␈↓ α←␈↓make inexact statements about the order of a sequence of events.␈↓
9␈↓

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[9]␈α∪By␈α∪phrasing␈α∩the␈α∪rule␈α∪correctly,␈α∩␈↓	DOBEFORE␈↓␈α∪can␈α∪be␈α∩used␈α∪to␈α∪state␈α∩four
␈↓ α←␈↓␈↓208    STRATEGIES␈↓ 
#7-4␈↓

␈↓"β␈↓ α←␈↓␈↓ β?When␈α∞all␈α∞the␈α∞meta-rules␈α∞have␈α∞been␈α∞applied,␈α∞they␈α∞will␈α∞have␈α∞made␈α∞a
␈↓ α←␈↓number␈αof␈αconclusions␈αabout␈αthe␈αutility␈αand␈αrelative␈αorder␈αof␈αthe␈αrules␈α
in␈αL.
␈↓ α←␈↓The␈α∞task␈α∞now␈α∞is␈α∞to␈α∞sort␈α∂and␈α∞perhaps␈α∞prune␈α∞L,␈α∞based␈α∞on␈α∂those␈α∞conclusions.
␈↓ α←␈↓Since␈α∞the␈α∞transitivity␈α∞of␈α∂the␈α∞order␈α∞relation␈α∞often␈α∞introduces␈α∂constraints␈α∞that
␈↓ α←␈↓are␈αnot␈αexplicitly␈αmentioned␈αby␈αa␈αmeta-rule,␈↓
10␈↓␈αit␈αis␈αfirst␈αnecessary␈αto␈αcompute
␈↓ α←␈↓the␈α∂transitive␈α∞closure␈α∂of␈α∂the␈α∞set␈α∂of␈α∞ordering␈α∂constraints.␈α∂ A␈α∞straightforward
␈↓ α←␈↓implementation of Warshall's algorithm [Warshall62] supplies this.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
pruning␈α
of␈α
L␈αis␈α
prompted␈α
by␈α
conclusions␈αlike␈α
those␈α
in␈α
Fig.␈α7-1,
␈↓ α←␈↓which␈α∂indicate␈α∂that␈α∂some␈α∂rules␈α∂will␈α∞definitely␈α∂be␈α∂useless␈α∂and␈α∂hence␈α∂can␈α∞be
␈↓ α←␈↓deleted. For the remainder, the most useful rules should be tried first.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂final␈α⊂step␈α⊂is␈α⊂thus␈α∂a␈α⊂sort-and-delete␈α⊂pass␈α⊂through␈α⊂L␈α⊂using␈α∂the
␈↓ α←␈↓following criteria:

␈↓"β␈↓ α←␈↓	If the utility of a rule is -1, delete it from L, otherwise
␈↓"β␈↓ α←␈↓	rule X goes before rule Y if
␈↓"β␈↓ α←␈↓	              it is required by ordering constraints, or
␈↓"β␈↓ α←␈↓	              the utility of X is higher than the utility of Y

␈↓ α←␈↓The result is a reordered and possibly shortened list.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈α
that␈α
with␈α
exhaustive␈αsearch␈α
of␈α
the␈α
goal␈αtree,␈α
order␈α
will␈α
make␈αa
␈↓ α←␈↓difference␈αin␈αthe␈αfinal␈αanswer␈αonly␈αif␈αthe␈αsystem␈αhappens␈αto␈αencounter␈αa␈αrule
␈↓ α←␈↓with␈α⊗ CF = 1.0 ␈α⊗that␈α⊗executes␈α⊗successfully,␈α⊗since␈α⊗in␈α⊗that␈α⊗case␈α↔search␈α⊗is
␈↓ α←␈↓terminated.␈α
 Even␈α
if␈α
the␈α
final␈α
answer␈α
is␈α
unchanged,␈α
however,␈α
the␈αprogram's
␈↓ α←␈↓performance␈α∂may␈α∂appear␈α∞more␈α∂rational␈α∂as␈α∂a␈α∞result␈α∂of␈α∂reordering␈α∂the␈α∞rules,
␈↓ α←␈↓since␈αit␈αwill␈αthen␈αtry␈αthe␈αmore␈α``appropriate''␈αlines␈αof␈αreasoning␈α
first.␈α Section
␈↓ α←␈↓7-4-6␈α∞considers␈α∂the␈α∞potential␈α∂impact␈α∞of␈α∞this␈α∂form␈α∞of␈α∂meta-rule␈α∞if␈α∂search␈α∞is
␈↓ α←␈↓not exhaustive.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈αalso␈αthat␈αeven␈αthough␈αutility␈αand␈αrule␈αorder␈αare␈αboth␈αeventually
␈↓ α←␈↓used␈α∞to␈α∞sort␈α∞the␈α
list,␈α∞they␈α∞are␈α∞maintained␈α
as␈α∞independent␈α∞factors,␈α∞since␈α
they
␈↓ α←␈↓represent␈α⊃two␈α⊃different␈α⊃kinds␈α⊃of␈α⊂judgments.␈α⊃ To␈α⊃see␈α⊃this,␈α⊃consider␈α⊃that␈α⊂it
␈↓ α←␈↓might,␈α∩for␈α⊃example,␈α∩be␈α⊃possible␈α∩to␈α⊃conclude␈α∩that␈α⊃two␈α∩rules␈α∩are␈α⊃definitely
␈↓ α←␈↓(CF = 1)␈α
going␈α
to␈α
be␈α∞especially␈α
useful;␈α
yet␈α
independent␈α∞considerations␈α
might
␈↓ α←␈↓still indicate that one of them should be invoked before the other.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αentire␈αprocess␈αis␈αsummarized␈αbelow␈α(simplified␈αby␈αassuming␈αthat
␈↓ α←␈↓there is only one level of meta-rules):




␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓different␈αrelationships: ␈α``Do␈αlist␈αX␈αbefore␈α
list␈αY''␈αand␈α``do␈αlist␈αX␈αafter␈α
list␈αY''
␈↓ α←␈↓are␈α
handled␈αin␈α
the␈α
obvious␈αway;␈α
``do␈α
X␈αlast''␈α
means␈α
``do␈αcomplement-X␈α
before
␈↓ α←␈↓X,'' and ``do X first'' means ``do X before complement-X.''

␈↓"β␈↓ α←␈↓[10]␈α
For␈αinstance,␈α
if␈α
``do␈αX␈α
before␈αY''␈α
and␈α
``do␈αY␈α
before␈αZ''␈α
are␈α
indicated␈αby
␈↓ α←␈↓rules,␈α
often␈α∞there␈α
is␈α∞no␈α
rule␈α∞that␈α
indicates␈α∞the␈α
necessary␈α∞condition␈α
of␈α∞``do␈α
X
␈↓ α←␈↓before Z.''
␈↓"β␈↓ α←␈↓␈↓7-4␈↓ λbMETA-RULES    209␈↓


␈↓"β␈↓ α←␈↓	 To deduce the value of an attribute A:
␈↓"β␈↓ α←␈↓	  1) L  ← the list of rules which conclude about A
␈↓"β␈↓ α←␈↓	  2) L' ← the list of meta-rules associated with A
␈↓"β␈↓ α←␈↓	  3) Evaluate each of the rules in L'; this may result in some
␈↓"β␈↓ α←␈↓	        conclusions about the rules in L
␈↓"β␈↓ α←␈↓	  4) Sort and prune L according to the criteria shown above
␈↓"β␈↓ α←␈↓	  5) Evaluate each of the rules in L; this may result in
␈↓"β␈↓ α←␈↓	        conclusions about the value of A


␈↓"β␈↓ α←␈↓␈↓α7-4-3    Implications of meta-rules as a strategy encoding␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈αare␈αseveral␈αadvantages␈αto␈αthis␈αapproach␈αto␈αencoding␈αstrategies.
␈↓ α←␈↓To␈αmake␈αthese␈αadvantages␈αclear,␈αrecall␈αthat␈αthe␈αbasic␈αcontrol␈αstructure␈αof␈αthe
␈↓ α←␈↓performance␈αprogram␈αis␈α
a␈αdepth-first␈αsearch␈αof␈α
the␈αand/or␈αgoal␈αtree␈α
sprouted
␈↓ α←␈↓by unwinding rules.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂first␈α⊂advantage␈α∂arises␈α∂from␈α⊂the␈α∂significant␈α⊂leverage␈α∂apparently
␈↓ α←␈↓available␈α∞from␈α
the␈α∞addition␈α
of␈α∞a␈α
store␈α∞of␈α
(meta-level)␈α∞knowledge␈α
describing
␈↓ α←␈↓which␈αchunk␈αof␈αobject-level␈αknowledge␈αto␈αinvoke␈αnext.␈α Considered␈αagain␈αin
␈↓ α←␈↓tree␈αsearch␈αterms,␈αwe␈αare␈αtalking␈αabout␈αthe␈αdifference␈αbetween␈α``blind''␈αsearch
␈↓ α←␈↓of␈αthe␈αtree␈α
and␈αone␈αguided␈αby␈α
heuristics.␈α The␈αadvantage␈α
of␈αeven␈αa␈αfew␈α
good
␈↓ α←␈↓heuristics␈α
in␈α
cutting␈αdown␈α
the␈α
combinatorial␈α
explosion␈αof␈α
tree␈α
search␈α
is␈αwell
␈↓ α←␈↓known.
␈↓"β␈↓ α←␈↓␈↓ β?Consider,␈α∀too,␈α∪that␈α∀part␈α∪of␈α∀the␈α∪definition␈α∀of␈α∀intelligence␈α∪includes
␈↓ α←␈↓appropriate␈αuse␈αof␈αinformation.␈α Even␈αif␈αa␈αstore␈αof␈α(object-level)␈αinformation
␈↓ α←␈↓is␈αnot␈αlarge,␈αit␈αis␈αimportant␈αto␈αbe␈αable␈αto␈αuse␈αit␈αproperly.␈α Meta-rules␈αprovide
␈↓ α←␈↓a␈α
mechanism␈α
for␈αencoding␈α
strategies␈α
that␈α
can␈αoffer␈α
additional␈α
guidance␈αto␈α
the
␈↓ α←␈↓system.
␈↓"β␈↓ α←␈↓␈↓ β?Third,␈α≤the␈α≤presence␈α≥of␈α≤meta-rules␈α≤associated␈α≥with␈α≤individual
␈↓ α←␈↓attributes␈α∞(as␈α∞the␈α∞rule␈α∞in␈α∞Fig.␈α∞7-1␈α∞is␈α∞associated␈α∞with␈α∞the␈α∂attribute␈α∞␈↓	IDENT␈↓ity)
␈↓ α←␈↓means␈α⊂that␈α⊂this␈α⊂goal␈α⊂tree␈α⊂has␈α⊂an␈α⊂interesting␈α⊂characteristic: ␈α⊂At␈α⊃each␈α⊂node,
␈↓ α←␈↓when␈α⊃the␈α⊃system␈α∩has␈α⊃to␈α⊃choose␈α⊃a␈α∩path,␈α⊃there␈α⊃may␈α⊃be␈α∩information␈α⊃stored
␈↓ α←␈↓advising␈α
about␈α
the␈α
best␈α
path␈α
to␈α
take.␈α
 There␈α
may␈α
therefore␈α
be␈α
available␈α
an
␈↓ α←␈↓extensive␈α
body␈α
of␈α
knowledge␈α
to␈α
guide␈α
the␈α
search,␈α
but␈α
this␈α
knowledge␈α∞is␈α
not
␈↓ α←␈↓embedded␈α
in␈α
the␈α∞code␈α
of␈α
a␈α
clever␈α∞search␈α
algorithm.␈α
 It␈α
is␈α∞instead␈α
organized
␈↓ α←␈↓around␈αthe␈αspecific␈αobjects␈αthat␈αform␈αthe␈αnodes␈αin␈αthe␈αtree;␈αthat␈αis,␈αinstead␈αof
␈↓ α←␈↓a smart algorithm, we have a ``smart tree.''
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
power␈α
in␈α
using␈αrules␈α
to␈α
guide␈α
rules␈αapplies␈α
at␈α
multiple␈α
levels␈αas
␈↓ α←␈↓well: ␈α⊗There␈α↔is␈α⊗leverage␈α⊗in␈α↔encoding␈α⊗heuristics␈α⊗that␈α↔guide␈α⊗the␈α↔use␈α⊗of
␈↓ α←␈↓heuristics.␈α  Thus,␈α∨rather␈α than␈α∨adding␈α more␈α∨heuristics␈α to␈α∨improve
␈↓ α←␈↓performance,␈α∀we␈α∀might␈α∀add␈α∃more␈α∀information␈α∀at␈α∀the␈α∀next␈α∃higher␈α∀level
␈↓ α←␈↓concerning the effective use of existing heuristics.
␈↓"β␈↓ α←␈↓␈↓ β?Fourth,␈αnote␈α
that␈αthe␈α
rules␈αcan␈α
be␈αjudgmental.␈α
 This␈αmakes␈αit␈α
possible
␈↓ α←␈↓to␈α
write␈αrules␈α
which␈αmake␈α
different␈α
conclusions␈αabout␈α
the␈αbest␈α
strategy␈αto␈α
use,
␈↓ α←␈↓and␈αthen␈αallows␈αthe␈αunderlying␈αmodel␈αof␈αconfirmation␈αto␈αweigh␈αthe␈α
evidence.
␈↓ α←␈↓That␈αis,␈αthe␈αstrategies␈α
can␈α``argue''␈αabout␈αthe␈αbest␈α
rule␈αto␈αuse,␈αand␈αthe␈α
strategy
␈↓ α←␈↓␈↓210    STRATEGIES␈↓ 
#7-4␈↓

␈↓"β␈↓ α←␈↓that␈α∞presents␈α∂the␈α∞best␈α∂case␈α∞(as␈α∞judged␈α∂by␈α∞the␈α∂confirmation␈α∞model)␈α∂will␈α∞win
␈↓ α←␈↓out.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂judgmental␈α⊂character␈α⊂also␈α⊂allows␈α⊂the␈α⊂novel␈α⊂possibility␈α⊂of␈α∂both
␈↓ α←␈↓inexact␈α
and␈α
conflicting␈α
statements␈α∞concerning␈α
relative␈α
order.␈α
 We␈α∞might,␈α
for
␈↓ α←␈↓instance,␈α
have␈α
two␈α
meta-rules␈α
that␈αoffer␈α
different␈α
opinions␈α
about␈α
the␈αorder␈α
of
␈↓ α←␈↓two␈α
sorts␈α
of␈α
object-level␈α
rules,␈αindicating␈α
that␈α
there␈α
is␈α
evidence␈α
that␈α``subset␈α
X
␈↓ α←␈↓should␈α⊂probably␈α∂(.6)␈α⊂be␈α∂done␈α⊂before␈α⊂subset␈α∂Y''␈α⊂and␈α∂that␈α⊂``subset␈α⊂Y␈α∂should
␈↓ α←␈↓probably␈α
(.4)␈α
be␈αdone␈α
before␈α
subset␈α
X.'' ␈αOnce␈α
again,␈α
the␈α
underlying␈αmodel␈α
of
␈↓ α←␈↓confirmation will weigh the evidence and produce an answer.␈↓
11␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Next,␈α_there␈α_are␈α_several␈α↔advantages␈α_associated␈α_with␈α_the␈α_use␈α↔of
␈↓ α←␈↓strategies␈αthat␈αare␈αgoal-specific␈αand␈αare␈αembedded␈αin␈αa␈αrepresentation␈αthat␈αis
␈↓ α←␈↓the␈α
same␈α
as␈αthat␈α
of␈α
the␈αobject-level␈α
knowledge.␈α
 The␈αfact␈α
that␈α
strategies␈αare
␈↓ α←␈↓␈↓↓goal-specific␈↓,␈αfor␈αinstance,␈αmakes␈αit␈αpossible␈αto␈αspecify␈αquite␈αprecise␈αheuristics
␈↓ α←␈↓for␈αa␈αgiven␈αgoal,␈αwithout␈αimposing␈αany␈αoverhead␈αin␈αthe␈αsearch␈αfor␈αany␈αother
␈↓ α←␈↓goals.␈α That␈α
is,␈αthere␈α
may␈αbe␈αa␈α
number␈αof␈α
complex␈αheuristics␈α
describing␈αthe
␈↓ α←␈↓best␈αrules␈αto␈αuse␈α
for␈αa␈αparticular␈αgoal,␈α
but␈αthese␈αwill␈αcause␈α
no␈αcomputational
␈↓ α←␈↓overhead except in the search for that goal.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α∃the␈α∃use␈α∃of␈α∃a␈α∃␈↓↓uniform␈α∃encoding␈α∃of␈α∃knowledge␈↓␈α⊗makes␈α∃the
␈↓ α←␈↓treatment␈α∃of␈α∃all␈α∃levels␈α∃of␈α∃knowledge␈α∃the␈α∃same,␈α∃and␈α∃this␈α⊗offers␈α∃several
␈↓ α←␈↓advantages.␈α_ For␈α→example,␈α_our␈α→work␈α_on␈α→explanation␈α_(chapter␈α→3)␈α_and
␈↓ α←␈↓knowledge␈αacquisition␈α(chapter␈α
5)␈αfor␈αobject-level␈α
rules␈αcan,␈αas␈α
a␈αresult␈αof␈α
the
␈↓ α←␈↓uniform␈α⊃encoding,␈α⊂easily␈α⊃be␈α⊂extended␈α⊃to␈α⊂meta-rules␈α⊃as␈α⊂well.␈α⊃ The␈α⊃first␈α⊂of
␈↓ α←␈↓these␈α_(explanation)␈α_has␈α↔been␈α_done␈α_and␈α↔makes␈α_possible␈α_an␈α↔interesting
␈↓ α←␈↓capability: ␈α∞In␈α∂addition␈α∞to␈α∂being␈α∞able␈α∂to␈α∞display␈α∂the␈α∞object-level␈α∂rules␈α∞used
␈↓ α←␈↓during␈α
a␈α
consultation,␈α
the␈α
system␈α
can␈α
similarly␈α
display␈α
the␈αmeta-rules,␈α
thereby
␈↓ α←␈↓making␈α
visible␈α
the␈α
criteria␈α
it␈α
used␈α∞in␈α
``deciding␈α
how␈α
to␈α
do␈α
what␈α
it␈α∞did''␈α
(see
␈↓ α←␈↓Section␈α⊃7-4-5).␈α⊃ Knowledge␈α⊃in␈α⊃the␈α⊂strategies␈α⊃has␈α⊃become␈α⊃accessible␈α⊃to␈α⊂the
␈↓ α←␈↓rest of the system and can be displayed in the same fashion.
␈↓"β␈↓ α←␈↓␈↓ β?Additional␈α∞advantages␈α∞associated␈α∞with␈α∞making␈α∞strategies␈α∞explicit␈α
are
␈↓ α←␈↓described in Section 7-5-3.

␈↓"β␈↓ α←␈↓␈↓α7-4-4    Advanced issues␈↓
␈↓"β␈↓ α←␈↓␈↓ β?For␈α∞the␈α
sake␈α∞of␈α∞clarity,␈α
the␈α∞description␈α
of␈α∞meta-rule␈α∞operation␈α
given
␈↓ α←␈↓above␈α⊂omitted␈α⊂some␈α⊂details␈α⊂of␈α⊂knowledge␈α⊂organization.␈α⊂ These␈α⊂details␈α∂and
␈↓ α←␈↓their implications are examined below.
␈↓"β␈↓ α←␈↓␈↓ β?Experience␈αwith␈αthe␈α
meta-rule␈αconstruct␈αhas␈αalso␈α
indicated␈αthat␈αit␈αis␈α
a
␈↓ α←␈↓convenient␈αmechanism␈αfor␈αencoding␈αforms␈αof␈αknowledge␈αin␈αaddition␈αto␈αthose
␈↓ α←␈↓described␈α
above.␈α Two␈α
such␈αuses␈α
are␈αdiscussed␈α
below,␈αdemonstrating␈α
that␈αit␈α
is

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[11]␈αNote␈αthat␈αthere␈αis,␈α
in␈αgeneral,␈αno␈αlogical␈αcontradiction␈αin␈α
the␈αconcurrent
␈↓ α←␈↓existence␈α≤of␈α≠evidence␈α≤suggesting␈α≠both␈α≤orderings.␈α≠ Only␈α≤one␈α≤case␈α≠is
␈↓ α←␈↓contradictory:␈αIn␈αthe␈α
current␈αmodel␈αit␈αwould␈α
be␈αcontradictory␈αto␈αhave␈α
definite
␈↓ α←␈↓evidence␈α↔(CF␈α↔=␈α_1)␈α↔for␈α↔both␈α↔the␈α_``before''␈α↔and␈α↔``after''␈α_orderings.␈α↔The
␈↓ α←␈↓implementation␈αof␈αmeta-rules␈αdoes␈αnot␈αnow␈αcheck␈αfor␈αthis␈αcase,␈αbut␈αit␈αshould
␈↓ α←␈↓someday be upgraded to do so.
␈↓ α←␈↓␈↓7-4␈↓ λbMETA-RULES    211␈↓

␈↓"β␈↓ α←␈↓possible␈αto␈αembody␈αin␈αmeta-rules␈αa␈αlimited␈αsubset␈αof␈αthe␈αknowledge␈αformerly
␈↓ α←␈↓embedded␈α
(sometimes␈αquite␈α
subtly)␈αin␈α
the␈αperformance␈α
program.␈α
 These␈αtwo
␈↓ α←␈↓types␈αof␈αrules␈αeither␈αcapture␈αaspects␈αof␈αthe␈αcontrol␈αstructure␈αor␈α
make␈αexplicit
␈↓ α←␈↓what we refer to as ``design decisions.''
␈↓"β␈↓ α←␈↓␈↓ β?Some␈α
of␈α
the␈α
material␈α
presented␈αhere␈α
relies␈α
on␈α
some␈α
fairly␈αspecific␈α
(and
␈↓ α←␈↓occasionally␈α"subtle)␈α"aspects␈α"of␈α"the␈α"present␈α#performance␈α"program's
␈↓ α←␈↓organization.␈α
 It␈α
is␈α
thus␈α
not␈α∞as␈α
widely␈α
relevant␈α
as␈α
preceding␈α
material␈α∞and␈α
is
␈↓ α←␈↓susceptible to change as experience with it increases.

␈↓"β␈↓ α←␈↓␈↓αMeta-rule organization:  Static␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?It␈α⊃was␈α⊂indicated␈α⊃earlier␈α⊂that␈α⊃meta-rules,␈α⊂like␈α⊃object-level␈α⊃rules,␈α⊂are
␈↓ α←␈↓associated␈αwith␈αspecific␈αattributes.␈α It␈αis␈αalso␈αpossible␈αto␈αassociate␈αa␈αmeta-rule
␈↓ α←␈↓with␈α∂any␈α⊂of␈α∂the␈α∂objects␈α⊂(e.g.,␈α∂an␈α∂infection,␈α⊂culture,␈α∂organism,␈α∂etc.).␈α⊂ In␈α∂that
␈↓ α←␈↓case,␈αthe␈αrule␈αis␈αused␈αas␈αa␈αstrategy␈αfor␈αall␈αattributes␈αof␈αthat␈αobject␈αand␈αfor␈αall
␈↓ α←␈↓attributes␈α∃of␈α∃any␈α∃object␈α∃that␈α∃gets␈α∃sprouted␈α∃below␈α∃it.␈↓
12␈↓␈α∃This␈α∃offers␈α∀the
␈↓ α←␈↓opportunity␈αto␈α
state␈αstrategies␈αwhose␈α
range␈αof␈αapplicability␈α
runs␈αfrom␈αa␈α
single
␈↓ α←␈↓object␈α∞(if␈α∞associated␈α∞with␈α∞a␈α∞leaf␈α∞of␈α
the␈α∞object␈α∞tree)␈α∞to␈α∞the␈α∞entire␈α∞domain␈α
(if
␈↓ α←␈↓associated␈α∃with␈α∀the␈α∃root␈α∃of␈α∀the␈α∃tree--domain-independent␈α∃strategies␈α∀are
␈↓ α←␈↓stored␈αthere).␈α The␈αspecificity␈αof␈αapplication␈αof␈αa␈αstrategy␈αis␈αthus␈αcontrollable
␈↓ α←␈↓by choosing the proper node in the object tree.

␈↓"β␈↓ α←␈↓␈↓αMeta-rule organization:  Dynamic␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Previous␈αexamples␈αhave␈αillustrated␈αtwo␈αdifferent␈αforms␈αof␈αmeta-rules.
␈↓ α←␈↓The␈α∞first␈α∞meta-rule␈α∂in␈α∞Fig.␈α∞7-1␈α∞concluded␈α∂about␈α∞the␈α∞␈↓↓individual␈↓␈α∞utility␈α∂of␈α∞a
␈↓ α←␈↓single␈αobject-level␈αrule,␈αwhile␈αthe␈αsecond␈αmeta-rule␈αdescribed␈αthe␈α␈↓↓comparative␈↓
␈↓ α←␈↓utility␈αof␈αtwo␈αclasses␈αof␈αrules.␈α
 In␈αterms␈αof␈αthe␈αtree␈αsearch␈αprocess␈α
that␈αmeta-
␈↓ α←␈↓rules␈α
guide,␈α
these␈α
two␈α
examples␈α
deal␈α
with␈α
different␈α
criteria␈α
for␈α
ordering␈αthe
␈↓ α←␈↓paths␈α⊂to␈α⊂be␈α⊂explored,␈α⊂based␈α⊂on␈α⊂a␈α⊂``look␈α⊂ahead''␈α⊂one␈α⊂level␈α⊂deep.␈α⊂ It␈α⊂is␈α∂also
␈↓ α←␈↓possible␈αto␈αexpress␈αsome␈αvery␈αsimple␈αstrategies␈αthat␈αuse␈αa␈αlook-ahead␈αseveral
␈↓ α←␈↓levels deep, to select a particular ``line of reasoning.''
␈↓"β␈↓ α←␈↓␈↓ β?A␈αstandard␈αproblem␈αin␈αdoing␈αthis␈αis␈αthe␈αquestion␈αof␈αhow␈αdeep␈αin␈αthe
␈↓ α←␈↓line␈αof␈αreasoning␈αto␈αexplore␈αor,␈αin␈αour␈αterms,␈αhow␈αlong␈αa␈αsequence␈αof␈αrules␈αto
␈↓ α←␈↓consider.␈α The␈αlonger␈αthe␈αsequence,␈αthe␈α
more␈αeffective␈αthe␈αchoice␈αmay␈αbe;␈α
but
␈↓ α←␈↓more␈α∞work␈α∞is␈α∞involved␈α∞in␈α∞simulating␈α∞the␈α∞rule␈α∞retrieval␈α∞mechanism␈α∞to␈α
effect
␈↓ α←␈↓the look-ahead.
␈↓"β␈↓ α←␈↓␈↓ β?We␈αhave␈αsidestepped␈αthe␈αissue␈αof␈αhow␈αdeep␈αto␈αlook␈α(and␈αsettled␈αfor␈αa
␈↓ α←␈↓weaker␈α∞solution)␈α∞by␈α∞implementing␈α∞a␈α∞``line-of-reasoning''␈α∞type␈α∞meta-rule␈α
that
␈↓ α←␈↓functions␈α∂in␈α∂a␈α∂somewhat␈α∂backward␈α∂fashion.␈α∂ Each␈α∂time␈α∂the␈α∂system␈α⊂tries␈α∂to
␈↓ α←␈↓establish␈α∀the␈α∪value␈α∀of␈α∪an␈α∀attribute,␈α∪it␈α∀retrieves␈α∪not␈α∀only␈α∀the␈α∪meta-rules
␈↓ α←␈↓associated␈αwith␈αthat␈αattribute␈αbut␈αany␈αassociated␈αwith␈αany␈αattribute␈αhigher␈αin
␈↓ α←␈↓the␈α∂current␈α∂chain␈α∂of␈α∂goals.␈α⊂ Thus,␈α∂instead␈α∂of␈α∂looking␈α∂deeper␈α∂from␈α⊂a␈α∂node,
␈↓ α←␈↓each␈αnode␈α``looks␈αup.''  ␈αThe␈αcombined␈αset␈αof␈αmeta-rules␈αis␈αused␈αto␈αreorder␈αor
␈↓ α←␈↓prune the current list of object rules.
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[12]␈αRecall␈αthat␈αobjects␈αare␈αorganized␈αinto␈αa␈αtree␈αof␈αthe␈αsort␈αshown␈αin␈αFig.␈α2-
␈↓ α←␈↓6.
␈↓ α←␈↓␈↓212    STRATEGIES␈↓ 
#7-4␈↓

␈↓"β␈↓ α←␈↓␈↓ β?In␈α≠simpler␈α≠terms: ␈α≠The␈α≠current␈α≠implementation␈α≠is␈α~functionally
␈↓ α←␈↓equivalent␈α⊂to␈α⊂(but␈α⊂more␈α⊂efficient␈α⊂than)␈α⊂taking␈α⊂each␈α⊂line-of-reasoning␈α⊂type
␈↓ α←␈↓meta-rule␈αand␈αputting␈αcopies␈αat␈αeach␈αnode␈αof␈αthe␈αsubtree␈αto␈αwhich␈αit␈αapplies.
␈↓ α←␈↓Thus,␈α⊂instead␈α⊂of␈α⊃having␈α⊂the␈α⊂system␈α⊂actively␈α⊃search␈α⊂ahead␈α⊂more␈α⊃than␈α⊂one
␈↓ α←␈↓level,␈α⊂the␈α⊃relevant␈α⊂strategy␈α⊃is␈α⊂``brought␈α⊃down''␈α⊂the␈α⊃tree␈α⊂during␈α⊃the␈α⊂normal
␈↓ α←␈↓search process.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
solution␈αis␈α
weaker␈αbecause␈α
it␈αis␈α
``hardwired''␈αinto␈α
the␈αdepth-first
␈↓ α←␈↓search␈α∞used␈α∂by␈α∞the␈α∞performance␈α∂program␈α∞and␈α∞because␈α∂it␈α∞only␈α∂searches␈α∞one
␈↓ α←␈↓level␈α∞ahead.␈α∞ A␈α∞true␈α∞line-of-reasoning␈α∞strategy␈α∞would␈α∞require␈α∞the␈α∂ability␈α∞to
␈↓ α←␈↓do its own examination of the search tree to an arbitrary number of levels.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α
are␈α
thus␈α
several␈α
potential␈α
sources␈α
of␈α
meta-rules␈α
for␈α
any␈αgiven
␈↓ α←␈↓attribute. They may be associated with:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?the attribute itself,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?the object to which that attribute applies,
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?any ancestor of that object, or
␈↓"β␈↓ α←␈↓␈↓ ββ(d)␈↓ β?any other attribute higher in the current chain of goals.

␈↓ α←␈↓Note␈α∞that␈α∂the␈α∞scope␈α∂of␈α∞the␈α∂first␈α∞three␈α∞of␈α∂these␈α∞is␈α∂fixed␈α∞but␈α∂that␈α∞the␈α∂last␈α∞is
␈↓ α←␈↓established dynamically as the consultation proceeds.

␈↓"β␈↓ α←␈↓␈↓αEncoding control structure information␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∩are␈α⊃object-level␈α∩rules␈α∩in␈α⊃the␈α∩system␈α⊃that␈α∩mention␈α∩the␈α⊃same
␈↓ α←␈↓attribute␈α_in␈α_their␈α→premise␈α_as␈α_they␈α_do␈α→in␈α_their␈α_actions.␈α→For␈α_example
␈↓ α←␈↓(paraphrasing): ␈α␈↓↓If␈α
you␈αthink␈α
the␈αidentity␈α
of␈αthe␈α
organism␈αis␈αpseudomonas,␈α
and
␈↓ α←␈↓↓the␈α∂patient␈α⊂has␈α∂skin␈α∂lesions,␈α⊂then␈α∂that's␈α∂strong␈α⊂additional␈α∂evidence␈α⊂that␈α∂the
␈↓ α←␈↓↓identity␈α_is␈α_in␈α↔fact␈α_pseudomonas.␈↓␈α_Such␈α_rules␈α↔are␈α_referred␈α_to␈α_as␈α↔``self-
␈↓ α←␈↓referencing.'' ␈α∞The␈α∞current␈α∞certainty␈α∞factor␈α∞model␈α∞requires␈α∞that␈α∞all␈α∞the␈α∞non-
␈↓ α←␈↓self-referencing␈α∩rules␈α⊃be␈α∩invoked␈α⊃before␈α∩those␈α⊃which␈α∩are␈α⊃self-referencing.
␈↓ α←␈↓(Failure␈α∩to␈α∩do␈α∩so␈α∩would␈α∩eliminate␈α∩the␈α∩commutative␈α∩property␈α∪of␈α∩certainty
␈↓ α←␈↓factors,␈αand␈αthe␈αfinal␈αresult␈αmight␈αbe␈αdependent␈αon␈αthe␈αorder␈αin␈αwhich␈αrules
␈↓ α←␈↓were invoked.)
␈↓"β␈↓ α←␈↓␈↓ β?In␈αvery␈αearly␈αversions␈αof␈αthe␈αsystem,␈αthe␈αmechanism␈αthat␈αinsured␈αthis
␈↓ α←␈↓partial␈αordering␈αwas␈αa␈αheavily␈αrecursive,␈αrather␈αobscure␈αpiece␈αof␈αcode.␈αIt␈αwas
␈↓ α←␈↓not␈α
at␈α
all␈α
obvious␈αwhat␈α
its␈α
purpose␈α
was,␈α
nor␈αthat␈α
this␈α
restriction␈α
on␈αrule␈α
order
␈↓ α←␈↓existed.␈α≤ A␈α≤later␈α≤version␈α≥included␈α≤a␈α≤separate␈α≤function␈α≤that␈α≥did␈α≤a
␈↓ α←␈↓straightforward␈α
reordering␈α
of␈α
the␈α
rule␈α
list␈α
just␈α
before␈α
it␈α
was␈α
invoked.␈α
 This␈α
at
␈↓ α←␈↓least␈αmade␈αfairly␈αobvious␈αwhat␈αwas␈αhappening␈αand␈αindicated␈αexplicitly␈αwhat
␈↓ α←␈↓the␈αordering␈α
constraint␈αwas.␈α
 Embedding␈αthis␈αconstraint␈α
in␈αa␈α
meta-rule␈α(Fig.
␈↓ α←␈↓7-2)␈α→represents␈α→a␈α→step␈α→toward␈α→a␈α→totally␈α→explicit,␈α→accessible␈α→chunk␈α_of
␈↓ α←␈↓knowledge.␈α
 The␈α
constraint␈α
is␈α
now␈α
quite␈α
clear,␈α
can␈α
be␈α
explained␈α
to␈α
the␈α
user
␈↓ α←␈↓by␈αthe␈αsystem␈αitself␈α(using␈αthe␈αexplanation␈αcapabilities␈αoutlined␈αin␈αchapter␈α3),
␈↓ α←␈↓and can be modified (if necessary) by editing the rule.␈↓
13␈↓

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[13]␈α⊗There␈α↔was␈α⊗actually␈α⊗an␈α↔intermediate␈α⊗solution␈α⊗which␈α↔illustrates␈α⊗the
␈↓ α←␈↓␈↓7-4␈↓ λbMETA-RULES    213␈↓


␈↓"β␈↓ α←␈↓	␈↓&METARULE003␈↓)αβ

␈↓"β␈↓ α←␈↓	If  1) there are rules which do not mention the current goal in
␈↓"β␈↓ α←␈↓	       their premise
␈↓"β␈↓ α←␈↓	    2) there are rules which mention the current goal in their
␈↓"β␈↓ α←␈↓	       premise
␈↓"β␈↓ α←␈↓	Then it is definite that the former should be done before the
␈↓"β␈↓ α←␈↓	latter.


␈↓"β␈↓ α←␈↓	PREMISE: ($AND(THEREARE OBJRULES($AND (DOESNTMENTION FREEVAR
␈↓"β␈↓ α←␈↓	                                       ACTION CURGOAL)) SET1)
␈↓"β␈↓ α←␈↓	              (THEREARE OBJRULES($AND (MENTIONS FREEVAR PREMISE
␈↓"β␈↓ α←␈↓	                                       CURGOAL) SET2))
␈↓"β␈↓ α←␈↓	ACTION:  (CONCLUDE SET1 DOBEFORE SET2 1000)


␈↓"β␈↓ α←␈↓α␈↓ βbFig. 7-2.    Example of a control structure meta-rule.    

␈↓ α←␈↓A␈α
part␈α
of␈α
the␈α
rule␈α
interpreter␈α
has␈α
itself␈α
been␈α
encoded␈α
in␈α
rule␈α
form.␈α∞ It␈α
may
␈↓ α←␈↓prove␈α∞possible␈α∞to␈α∞express␈α∞additional␈α∞parts␈α∞of␈α∞the␈α∞control␈α∞structure␈α∞in␈α
meta-
␈↓ α←␈↓rules.␈α Each␈αtime␈αthis␈αis␈αaccomplished␈αit␈αmeans␈αthat␈αsome␈α
additional␈αelement
␈↓ α←␈↓of␈α
the␈α
system's␈αbehavior␈α
becomes␈α
accessible␈α
and,␈αhence,␈α
can␈α
be␈α
explained␈αor
␈↓ α←␈↓modified using the existing facilities.

␈↓"β␈↓ α←␈↓␈↓αEncoding design decision information␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?For␈α∞the␈α∞sake␈α∞of␈α∞human␈α∂engineering,␈α∞it␈α∞makes␈α∞good␈α∞sense␈α∂during␈α∞an
␈↓ α←␈↓infectious␈α⊃disease␈α⊃consultation␈α⊃to␈α⊂ask␈α⊃about␈α⊃positive␈α⊃cultures␈α⊃(those␈α⊂which
␈↓ α←␈↓displayed␈αbacterial␈αgrowth)␈αbefore␈αthose␈αthat␈αturned␈αup␈αnegative.␈α
 Originally,
␈↓ α←␈↓this␈αdecision␈α
was␈αembedded␈αquite␈α
subtly␈αin␈α
the␈αordering␈αof␈α
a␈αlist␈α
internal␈αto
␈↓ α←␈↓the␈αsystem␈αcode.␈α It␈αcan␈αbe␈αstated␈αeasily␈αin␈αa␈αmeta-rule,␈αhowever,␈αas␈αshown␈αin
␈↓ α←␈↓Fig. 7-3.








␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓standard␈α∩conflict␈α∩between␈α∩efficiency␈α⊃and␈α∩comprehensibility.␈α∩ When␈α∩it␈α⊃was
␈↓ α←␈↓agreed␈α∩that␈α∩changes␈α∩to␈α∩the␈α∩knowledge␈α⊃base␈α∩would␈α∩not␈α∩be␈α∩made␈α∩while␈α⊃a
␈↓ α←␈↓consultation␈α
was␈α
in␈α
progress,␈α
it␈α
became␈α
possible␈α
to␈α
meet␈α
the␈α
constraint␈α
by␈α
pre-
␈↓ α←␈↓setting␈α∂the␈α∂order␈α∂of␈α∂each␈α∂internal␈α⊂list␈α∂of␈α∂rules␈α∂by␈α∂hand␈α∂and␈α⊂keeping␈α∂them
␈↓ α←␈↓stored␈α∞in␈α∞that␈α∞order.␈α∞ This␈α∞eliminates␈α∞all␈α∞execution␈α∞time␈α∞overhead,␈α∞but␈α∞also
␈↓ α←␈↓means␈α⊃that␈α⊃all␈α⊃representation␈α⊃of␈α⊃the␈α⊃constraint␈α⊃has␈α⊃disappeared␈α⊃from␈α⊃the
␈↓ α←␈↓system.
␈↓ α←␈↓␈↓214    STRATEGIES␈↓ 
#7-4␈↓


␈↓"β␈↓ α←␈↓	␈↓&METARULE004␈↓)αβ

␈↓"β␈↓ α←␈↓	If  1) there are rules which are relevant to positive cultures,
␈↓"β␈↓ α←␈↓	       and
␈↓"β␈↓ α←␈↓	    2) there are rules which are relevant to negative cultures
␈↓"β␈↓ α←␈↓	Then it is definite that the former should be done before the
␈↓"β␈↓ α←␈↓	latter.


␈↓"β␈↓ α←␈↓	PREMISE: ($AND(THEREARE OBJRULES ($AND (APPLIESTO FREEVAR
␈↓"β␈↓ α←␈↓	                                        POSCUL)) SET1)
␈↓"β␈↓ α←␈↓	              (THEREARE OBJRULES ($AND (APPLIESTO FREEVAR
␈↓"β␈↓ α←␈↓	                                        NEGCUL)) SET2))
␈↓"β␈↓ α←␈↓	ACTION:  (CONCLUDE SET1 DOBEFORE SET2 1000)


␈↓"β␈↓ α←␈↓α␈↓ βkFig. 7-3.    Example of a design decision meta-rule.    

␈↓"β␈↓ α←␈↓␈↓ β?Such␈α∂decisions,␈α∂based␈α∂on␈α∂human␈α∂engineering,␈α∂are␈α∂common␈α⊂in␈α∂user-
␈↓ α←␈↓oriented␈α∂systems.␈α⊂ Expressing␈α∂this␈α⊂information␈α∂in␈α∂a␈α⊂rule␈α∂renders␈α⊂it␈α∂explicit
␈↓ α←␈↓and accessible, making possible its explanation and modification.
␈↓"β␈↓ α←␈↓␈↓ β?In␈αsummary,␈αsince␈αmeta-rules␈αact␈αby␈αpruning␈αor␈αreordering␈αthe␈αlist␈αof
␈↓ α←␈↓rules␈α⊂to␈α⊂be␈α⊂invoked,␈α⊂they␈α⊂can␈α⊃be␈α⊂used␈α⊂to␈α⊂express␈α⊂any␈α⊂part␈α⊂of␈α⊃the␈α⊂control
␈↓ α←␈↓structure or the design decisions that can be understood in those terms.

␈↓"β␈↓ α←␈↓␈↓α7-4-5    Explanation and acquisition␈↓
␈↓"β␈↓ α←␈↓␈↓ β?One␈α
important␈α
advantage␈α
of␈α
a␈α
uniform␈α
encoding␈α
of␈α
different␈αlevels␈α
of
␈↓ α←␈↓knowledge␈α
is␈α
the␈αability␈α
to␈α
use␈α
a␈αsingle␈α
set␈α
of␈α
facilities␈αfor␈α
all␈α
levels.␈α
 In␈αthe
␈↓ α←␈↓case␈α⊃of␈α⊃explanation,␈α⊂for␈α⊃instance,␈α⊃all␈α⊂the␈α⊃machinery␈α⊃described␈α⊃earlier␈α⊂that
␈↓ α←␈↓deals␈α
with␈αobject-level␈α
rules␈αcan␈α
be␈α
employed␈αto␈α
deal␈αwith␈α
meta-rules.␈α To␈α
do
␈↓ α←␈↓this,␈α
it␈α
proved␈α
necessary␈α
to␈α
make␈α
only␈α
two␈α
small␈α
extensions,␈α
augmenting␈αthe
␈↓ α←␈↓existing␈α∞record␈α
of␈α∞rule␈α
invocations␈α∞to␈α
include␈α∞meta-rules␈α
and␈α∞providing␈α
the
␈↓ α←␈↓natural␈α∪language␈α∀facilities␈α∪for␈α∪explaining␈α∀the␈α∪meta-rule's␈α∀contribution␈α∪to
␈↓ α←␈↓program␈α
performance.␈α
 An␈α
example␈α
of␈α
the␈α
capability␈α
that␈α
resulted␈α
is␈αshown
␈↓ α←␈↓in␈α⊃a␈α⊂consultation␈α⊃segment␈α⊂below.␈α⊃ (Because␈α⊂␈↓	METARULE001␈↓␈α⊃refers␈α⊃to␈α⊂culture
␈↓ α←␈↓source,␈α∀which␈α∀is␈α∀already␈α∀known␈α∀when␈α∀the␈α∀rule␈α∀is␈α∀invoked,␈α∀it␈α∀does␈α∀not
␈↓ α←␈↓normally␈αgenerate␈αany␈αquestions␈αand␈αis␈αthus␈αtransparent␈αto␈αthe␈αuser.␈α For␈αthe
␈↓ α←␈↓sake␈α∪of␈α∪illustration,␈α∪the␈α∪first␈α∪clause␈α∪has␈α∪been␈α∪changed␈α∪to␈α∪one␈α∪that␈α∩does
␈↓ α←␈↓generate␈α≠a␈α~question,␈α≠giving␈α≠the␈α~user␈α≠an␈α~opportunity␈α≠to␈α≠request␈α~an
␈↓ α←␈↓explanation.)

␈↓"β␈↓ α←␈↓	--------ORGANISM-1--------
␈↓"β␈↓ α←␈↓	10) Enter the identity of ORGANISM-1:
␈↓"β␈↓ α←␈↓	**␈↓αUNK␈↓	
␈↓"β␈↓ α←␈↓	11) Is ORGANISM-1 a rod or coccus (etc.):
␈↓"β␈↓ α←␈↓	**␈↓αROD␈↓	
␈↓ α←␈↓␈↓7-4␈↓ λbMETA-RULES    215␈↓

␈↓"β␈↓ α←␈↓	12) The gram stain of ORGANISM-1:
␈↓"β␈↓ α←␈↓	**␈↓αNEG␈↓	
␈↓"β␈↓ α←␈↓	13) Do you consider George Filbert to be a compromised host
␈↓"β␈↓ α←␈↓	    (alcoholic, immunosuppressed, etc.)?
␈↓"β␈↓ α←␈↓	**␈↓αWHY␈↓	

␈↓"β␈↓ α←␈↓	[i.e. WHY is it important to determine whether or not PATIENT-4
␈↓"β␈↓ α←␈↓	 is a compromised host?]

␈↓"β␈↓ α←␈↓	[1.0] This will help determine a strategy for deducing the
␈↓"β␈↓ α←␈↓	      identity of ORGANISM-1.
␈↓"β␈↓ α←␈↓	      If
␈↓"β␈↓ α←␈↓	 [1.1] the patient is a compromised host, and
␈↓"β␈↓ α←␈↓	 [1.2] there are rules which mention in their premise a
␈↓"β␈↓ α←␈↓	       previous organism which may be the same as the current
␈↓"β␈↓ α←␈↓	       organism,
␈↓"β␈↓ α←␈↓	      then
␈↓"β␈↓ α←␈↓	 It is definite (1.0) that each of them is not going to be
␈↓"β␈↓ α←␈↓	 useful.
␈↓"β␈↓ α←␈↓	 [METARULE001]

␈↓"β␈↓ α←␈↓	**␈↓αHOW [1.0]␈↓	

␈↓"β␈↓ α←␈↓	[i.e. HOW was the strategy for deducing the identity of
␈↓"β␈↓ α←␈↓	 ORGANISM-1 determined?]

␈↓"β␈↓ α←␈↓	At that point METARULE001 was being used.
␈↓"β␈↓ α←␈↓	   In addition, the following were also used:
␈↓"β␈↓ α←␈↓	   [1.3] METARULE003 indicated it is definite that the former
␈↓"β␈↓ α←␈↓	           should be done before the latter

␈↓ α←␈↓↓␈↓ βW{Due␈αto␈αthe␈αprimitive␈αapproach␈αto␈αnatural␈αlanguage,␈αrule␈αtranslation
␈↓ α←␈↓↓␈↓ βWcan␈αbe␈αa␈αbit␈αcryptic␈αat␈αtimes.␈α However,␈αsee␈αbelow,␈αwhere␈αthe␈αfull␈αrule
␈↓ α←␈↓↓␈↓ βWtext makes things a little clearer.}

␈↓"β␈↓ α←␈↓	**␈↓αHOW [1.3]␈↓	

␈↓"β␈↓ α←␈↓	[i.e. HOW was METARULE003 used?]

␈↓"β␈↓ α←␈↓	If [2.1] there are rules which do not mention the current goal
␈↓"β␈↓ α←␈↓	         in their premise, and
␈↓"β␈↓ α←␈↓	   [2.2] there are rules which mention the current goal in
␈↓"β␈↓ α←␈↓	         their premise
␈↓"β␈↓ α←␈↓	Then it is definite that the former should be done before the
␈↓"β␈↓ α←␈↓	latter.
␈↓"β␈↓ α←␈↓	[METARULE003]

␈↓"β␈↓ α←␈↓␈↓ β?We␈αhave␈αnot␈αyet␈α
attempted␈αthe␈αinteractive␈αacquisition␈αof␈α
meta-rules--
␈↓ α←␈↓their␈α
slightly␈α
more␈α
complex␈α
syntax␈α
requires␈α
generalizing␈α
some␈α
of␈α∞the␈α
details
␈↓ α←␈↓of␈α⊂the␈α⊂acquisition␈α⊂routines.␈α⊂ It␈α⊂does␈α⊂not␈α⊂appear␈α⊂to␈α⊂be␈α⊂a␈α⊃difficult␈α⊂problem,
␈↓ α←␈↓however,␈α∞and␈α∞beyond␈α∞it␈α∞lie␈α∞some␈α∞interesting␈α∞issues.␈α∞ Given␈α∞the␈α∞success␈α∞with
␈↓"β␈↓ α←␈↓␈↓216    STRATEGIES␈↓ 
#7-4␈↓

␈↓"β␈↓ α←␈↓acquisition␈α∂of␈α⊂object-level␈α∂rules␈α∂in␈α⊂the␈α∂context␈α⊂of␈α∂a␈α∂bug,␈α⊂can␈α∂the␈α⊂same␈α∂be
␈↓ α←␈↓done␈αwith␈αmeta-rules?␈α What␈αare␈α
the␈αsuperficial␈αmanifestations␈αof␈αa␈αbug␈α
that
␈↓ α←␈↓suggest␈α⊃the␈α⊃need␈α⊃for␈α⊃a␈α⊃change␈α⊃in␈α⊃the␈α⊃way␈α⊃established␈α⊃knowledge␈α⊃is␈α⊃used
␈↓ α←␈↓(rather␈α⊃than␈α⊂a␈α⊃change␈α⊂in␈α⊃the␈α⊃content␈α⊂of␈α⊃the␈α⊂knowledge␈α⊃itself)?␈α⊃ We␈α⊂have
␈↓ α←␈↓considered␈αthis␈αbriefly␈αand␈αthe␈αmost␈αobvious␈αcase␈αseems␈αto␈αbe␈αone␈αanalogous
␈↓ α←␈↓to␈α∩human␈α∩behavior.␈α∩ When␈α∩someone␈α⊃inexperienced␈α∩in␈α∩a␈α∩domain␈α∩takes␈α⊃a
␈↓ α←␈↓plausible,␈α
but␈α
(to␈α
the␈α
expert)␈α
inappropriate␈α
approach␈α
to␈α
a␈α
problem,␈αthe␈α
expert
␈↓ α←␈↓can␈α∞often␈α∂spot␈α∞it␈α∂and␈α∞inquire␈α∞why␈α∂the␈α∞novice␈α∂took␈α∞that␈α∂approach.␈α∞ Similar
␈↓ α←␈↓situations␈α
may␈α
be␈α
encountered␈α
in␈αwhich␈α
the␈α
program␈α
starts␈α
off␈α
on␈αa␈α
plausible,
␈↓ α←␈↓but␈α∞incorrect␈α
track;␈α∞the␈α
appropriate␈α∞remedy␈α
may␈α∞be␈α
a␈α∞new␈α∞meta-rule.␈α
 This
␈↓ α←␈↓appears to be one interesting direction for future work.

␈↓"β␈↓ α←␈↓␈↓α7-4-6    Limitations of the current implementation, future work␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈α
are␈α
several␈α
shortcomings␈α
in␈α
the␈α
present␈α
implementation.␈α
Some
␈↓ α←␈↓are␈α∂due␈α∞to␈α∂fundamental␈α∞issues␈α∂of␈α∞system␈α∂organization,␈α∞while␈α∂others␈α∂are␈α∞the
␈↓ α←␈↓result␈α∞of␈α∞limited␈α∞experience␈α∞with␈α∂these␈α∞constructs␈α∞and,␈α∞thus,␈α∞may␈α∂change␈α∞in
␈↓ α←␈↓time.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α≤limitation␈α≤arises␈α≤from␈α≤the␈α≤way␈α≤meta-rules␈α≤augment␈α≠the
␈↓ α←␈↓performance␈α∀program's␈α∀control␈α∀structure.␈α∀ In␈α∀the␈α∀present␈α∪implementation,
␈↓ α←␈↓their␈α∞effect␈α
is␈α∞constrained␈α
to␈α∞the␈α
reordering␈α∞or␈α
pruning␈α∞of␈α
rule␈α∞lists.␈α
 While
␈↓ α←␈↓this␈αis␈α
useful,␈αexamples␈α
above␈αdemonstrate␈α
that␈αit␈α
offers␈αa␈α
limited␈αcapability
␈↓ α←␈↓for␈αencoding␈αaspects␈αof␈αthe␈αcontrol␈αstructure.␈α More␈αimportant,␈αit␈αprovides␈αno
␈↓ α←␈↓means␈α
of␈αeffecting␈α
the␈αmore␈α
interesting␈αcapability␈α
of␈αswitching␈α
methodologies.
␈↓ α←␈↓Since␈αmedical␈αdiagnosis␈αis␈αknown␈αto␈α
be␈αa␈αcomplex␈αset␈αof␈αbehaviors␈α
(see,␈αe.g.,
␈↓ α←␈↓[Miller75]),␈αit␈αmight␈αprove␈αuseful␈αto␈αbe␈αable␈αto␈αshift␈αback␈αand␈αforth␈αbetween,
␈↓ α←␈↓say,␈α→a␈α~goal-directed␈α→and␈α~a␈α→data-directed␈α→invocation␈α~of␈α→rules␈α~as␈α→the
␈↓ α←␈↓consultation␈α∂progresses.␈α∞ (See␈α∂Section␈α∞7-4-6␈α∂for␈α∞some␈α∂thoughts␈α∞on␈α∂how␈α∞this
␈↓ α←␈↓might be accomplished.)
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∂second␈α∞limitation␈α∂is␈α∞the␈α∂restriction␈α∞on␈α∂circumstances␈α∂under␈α∞which
␈↓ α←␈↓meta-rules␈α⊃are␈α⊃invoked.␈α⊃ As␈α⊃illustrated,␈α⊃they␈α⊃are␈α⊃invoked␈α⊃when␈α⊃a␈α∩goal␈α⊃is
␈↓ α←␈↓sought␈α⊂or␈α∂when␈α⊂a␈α∂conclusion␈α⊂is␈α⊂made.␈α∂ There␈α⊂are␈α∂many␈α⊂other␈α⊂events␈α∂that
␈↓ α←␈↓could␈α⊂quite␈α∂usefully␈α⊂trigger␈α∂them.␈α⊂ In␈α⊂particular,␈α∂the␈α⊂␈↓↓failure␈↓␈α∂of␈α⊂one␈α⊂of␈α∂the
␈↓ α←␈↓object-level␈α∩rules␈α∪might␈α∩be␈α∪very␈α∩informative␈α∩and␈α∪could␈α∩suggest␈α∪a␈α∩useful
␈↓ α←␈↓reordering␈α
or␈α
pruning␈α
of␈α
the␈α
remainder␈α
of␈α
the␈α
list.␈α
 The␈α
basic␈α
problem␈α
is␈α
that
␈↓ α←␈↓meta-rules␈α∀are␈α∀simply␈α∀not␈α∀a␈α∀general␈α∀demon-like␈α∀mechanism.␈α∃ This␈α∀may
␈↓ α←␈↓eventually␈α
be␈α
changed␈α
if␈α
a␈α
reasonably␈α
efficient␈α
scheme␈α
can␈α
be␈α∞found␈α
which
␈↓ α←␈↓will avoid some of the overhead typically associated with demons.
␈↓"β␈↓ α←␈↓␈↓ β?One␈αparticularly␈αuseful␈αapplication␈αof␈αsuch␈αa␈αscheme␈αwould␈αbe␈αits␈α
use
␈↓ α←␈↓in␈α
conjunction␈α
with␈α
non-exhaustive␈αapplication␈α
of␈α
the␈α
object-level␈αrules.␈α
 We
␈↓ α←␈↓noted␈α∃above␈α∃that,␈α∃due␈α∃to␈α∀the␈α∃exhaustive␈α∃search␈α∃currently␈α∃used␈α∃by␈α∀the
␈↓ α←␈↓performance␈α⊃program,␈α∩meta-rules␈α⊃typically␈α⊃do␈α∩not␈α⊃change␈α⊃the␈α∩final␈α⊃result
␈↓ α←␈↓reached␈α∂by␈α∂the␈α∂program␈α∂but␈α∂can,␈α∂instead,␈α∂make␈α∂the␈α∂program␈α⊂appear␈α∂more
␈↓ α←␈↓rational␈α
as␈α
it␈α
works␈α
to␈αdetermine␈α
the␈α
answer.␈α
 Non-exhaustive␈α
search␈αmight
␈↓ α←␈↓be␈α∞implemented␈α∂by␈α∞adding␈α∂``after-the-fact''␈α∞meta-rules␈α∂to␈α∞the␈α∂system:␈α∞These
␈↓ α←␈↓would␈αbe␈αmeta-rules␈α
executed␈αafter␈αan␈αattempt␈α
had␈αbeen␈αmade␈α
to␈αdetermine
␈↓ α←␈↓␈↓7-4␈↓ λbMETA-RULES    217␈↓

␈↓"β␈↓ α←␈↓the␈αvalue␈αof␈αa␈αspecific␈αattribute.␈α In␈αthis␈αcase␈αmeta-rules␈αexecuted␈αbefore␈αthe
␈↓ α←␈↓search␈α
might␈α
indicate␈α
that␈αonly␈α
one␈α
or␈α
two␈αreasoning␈α
paths␈α
should␈α
be␈αtried.
␈↓ α←␈↓Meta-rules␈α
executed␈α
after␈α
the␈α
search␈α
could␈α
test␈α
the␈α
strength␈α
of␈α
the␈α
conclusions
␈↓ α←␈↓made␈αso␈αfar␈αand␈αcould␈αperhaps␈αindicate␈αthat␈αthe␈αsystem␈αshould␈α``go␈αback␈αand
␈↓ α←␈↓try␈α∩again,''␈α∩this␈α∩time␈α∩exploring␈α∩additional␈α∩paths.␈α∩ This␈α∩would␈α∩permit␈α⊃the
␈↓ α←␈↓creation␈α
of␈α∞strategies␈α
that␈α∞decide␈α
when␈α
the␈α∞result␈α
is␈α∞``good␈α
enough''␈α∞to␈α
make
␈↓ α←␈↓further␈α
search␈α
unprofitable␈α
and␈α
may␈α
prove␈α
quite␈α
useful␈α
when␈α
the␈α
rule␈αbase
␈↓ α←␈↓becomes very large.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩present␈α∪implementation␈α∩of␈α∩meta-rules␈α∪also␈α∩has␈α∩a␈α∪problem␈α∩of
␈↓ α←␈↓comprehensibility: ␈α
The␈α
translations␈α
of␈α
the␈α
meta-rules␈α
shown␈α
earlier␈α
are␈αnot
␈↓ α←␈↓particularly␈α
lucid.␈α
 This␈α
is␈α
due␈αprimarily␈α
to␈α
the␈α
primitive␈α
approach␈αto␈α
natural
␈↓ α←␈↓language␈α∩used␈α∪in␈α∩the␈α∩system,␈α∪especially␈α∩the␈α∩clause-by-clause␈α∪approach␈α∩to
␈↓ α←␈↓translation␈α∀of␈α∀rules.␈α∀ As␈α∀the␈α∀meta-rules␈α∀begin␈α∀to␈α∀play␈α∀a␈α∀larger␈α∀part␈α∪in
␈↓ α←␈↓controlling␈αthe␈αsystem's␈αbehavior,␈αit␈αwill␈αbecome␈αimportant␈αto␈αprovide␈αa␈αmore
␈↓ α←␈↓sophisticated translation, one which yields more comprehensible results.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α⊂are␈α⊂also␈α⊂limitations␈α⊂that␈α⊂arise␈α⊂out␈α⊂of␈α⊂the␈α⊂decision␈α⊃to␈α⊂encode
␈↓ α←␈↓strategies␈α
in␈αthe␈α
same␈α
production␈αrule␈α
format␈α
as␈αthe␈α
object-level␈αrules.␈α
 While
␈↓ α←␈↓this␈α⊂offers␈α⊂the␈α⊂important␈α⊂advantage␈α⊂of␈α⊂a␈α⊂uniform␈α⊂encoding␈α⊂of␈α⊂knowledge,
␈↓ α←␈↓production␈α⊂rules␈α⊃are␈α⊂not␈α⊃well␈α⊂suited␈α⊃to␈α⊂the␈α⊃expression␈α⊂of␈α⊃complex␈α⊂control
␈↓ α←␈↓structures.␈α In␈αparticular,␈αthey␈α
are␈αnot␈αa␈αparticularly␈αtransparent␈α
medium␈αfor
␈↓ α←␈↓expressing␈α
a␈α
``chunk''␈αof␈α
behavior␈α
that␈αrequires␈α
more␈α
than␈αa␈α
single␈α
rule␈α(see
␈↓ α←␈↓[Davis77a]␈α∞for␈α∂a␈α∞discussion␈α∞of␈α∂this␈α∞point).␈α∞ More␈α∂generally,␈α∞the␈α∞issue␈α∂of␈α∞an
␈↓ α←␈↓appropriate strategy language is still an open question.
␈↓"β␈↓ α←␈↓␈↓ β?Problems␈α
also␈α
arise␈α
from␈α
the␈α
restricted␈α
syntax␈α
available␈α
in␈α
␈↓¬MYCIN␈↓␈αand
␈↓ α←␈↓adopted,␈αfor␈α
the␈αsake␈α
of␈αuniformity,␈αin␈α
␈↓¬TEIRESIAS␈↓.␈α In␈α
particular,␈αthe␈α
fact␈αthat
␈↓ α←␈↓the␈α∩only␈α⊃``action''␈α∩currently␈α⊃available␈α∩is␈α⊃to␈α∩make␈α⊃a␈α∩conclusion␈α∩means␈α⊃that
␈↓ α←␈↓meta-rules␈α∞can␈α∞be␈α∂used␈α∞to␈α∞perform␈α∂only␈α∞a␈α∞limited␈α∂``direction''␈α∞of␈α∞the␈α∂use␈α∞of
␈↓ α←␈↓object-level rules, reordering or pruning the list before it is invoked.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α
there␈α
is␈α
the␈α
subtlety␈α
of␈α
the␈α
distinction␈α
in␈α
the␈αimplementation
␈↓ α←␈↓of␈α⊃the␈α⊃``statement-of-utility''␈α⊃and␈α⊃``line-of-reasoning''␈α⊃types␈α⊃of␈α⊃strategy.␈α⊃ As
␈↓ α←␈↓indicated␈α
earlier,␈αat␈α
each␈αgoal␈α
the␈α
system␈αretrieves␈α
all␈αthe␈α
meta-rules␈α
for␈αthe
␈↓ α←␈↓current␈α∞attribute,␈α∂as␈α∞well␈α∞as␈α∂for␈α∞any␈α∂attributes␈α∞above␈α∞it␈α∂in␈α∞the␈α∂current␈α∞goal
␈↓ α←␈↓chain.␈α∂ It␈α∂is␈α∞necessary␈α∂to␈α∂distinguish␈α∞between␈α∂rules␈α∂that␈α∞are␈α∂intended␈α∂to␈α∞be
␈↓ α←␈↓applicable␈α∞at␈α∂only␈α∞a␈α∞single␈α∂level␈α∞(i.e.,␈α∞``statement␈α∂of␈α∞utility''␈α∞types)␈α∂and␈α∞those
␈↓ α←␈↓that␈α
apply␈αto␈α
an␈α
arbitrary␈αnumber␈α
of␈αlevels␈α
(i.e.,␈α
anywhere␈αalong␈α
the␈α
line␈αof
␈↓ α←␈↓reasoning␈α⊃leading␈α⊃to␈α⊃the␈α⊃goal).␈α⊃This␈α⊃is␈α⊃handled␈α⊃currently␈α⊃by␈α⊃a␈α⊂somewhat
␈↓ α←␈↓obscure␈α∀solution␈α∀to␈α∀what␈α∀is␈α∀really␈α∀a␈α∀much␈α∀larger,␈α∀nontrivial␈α∃issue: ␈α∀the
␈↓ α←␈↓problem␈α⊃of␈α∩implicit␈α⊃context,␈α∩described␈α⊃in␈α⊃Section␈α∩2-4.␈α⊃ The␈α∩difference␈α⊃in
␈↓ α←␈↓effect␈αof␈αassociating␈αa␈αmeta-rule␈αwith␈αan␈αattribute␈αas␈αopposed␈αto␈αan␈αobject␈αis
␈↓ α←␈↓another aspect of this same issue and suffers from a similar obscurity.
␈↓ α←␈↓␈↓218    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓␈↓α7-5    BROADER IMPLICATIONS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
discussion␈α
thus␈αfar␈α
has␈α
centered␈αaround␈α
the␈α
concept␈αof␈α
saturation
␈↓ α←␈↓and␈αour␈αmethod␈αfor␈αdealing␈αwith␈αit,␈αnamely,␈αthe␈αuse␈αof␈αmeta-level␈αknowledge
␈↓ α←␈↓as␈α∀a␈α∀site␈α∀for␈α∀embedding␈α∀strategies␈α∀to␈α∀guide␈α∀the␈α∀program.␈α∃ As␈α∀previous
␈↓ α←␈↓examples␈α⊃have␈α∩shown,␈α⊃this␈α∩technique␈α⊃is␈α⊃also␈α∩useful␈α⊃for␈α∩guiding␈α⊃program
␈↓ α←␈↓performance even when the system is not saturated.
␈↓"β␈↓ α←␈↓␈↓ β?At␈α⊃this␈α⊃point␈α⊃the␈α⊃discussion␈α⊃broadens␈α⊃to␈α⊃consider␈α⊃some␈α⊂interesting
␈↓ α←␈↓additional␈α_issues␈α→that␈α_arise␈α→from␈α_examining␈α→the␈α_mechanisms␈α→used␈α_to
␈↓ α←␈↓implement␈αmeta-rules.␈α We␈α
consider␈αin␈αparticular␈α
the␈αimplications␈αthat␈α
follow
␈↓ α←␈↓from␈α
two␈αfeatures: ␈α
(a)  meta-rules'␈αability␈α
to␈α
select␈αobject-level␈α
rules␈αby␈α
direct
␈↓ α←␈↓examination␈α
of␈αobject-level␈α
rule␈α
code␈αand␈α
(b)  meta-rules'␈αuse␈α
of␈α
an␈αexplicit
␈↓ α←␈↓functional␈α∪specification␈α∩of␈α∪retrieval␈α∪criteria.␈α∩ The␈α∪focus␈α∩here␈α∪is␈α∪on␈α∩these
␈↓ α←␈↓techniques␈αas␈α
they␈αapply␈α
to␈αprogramming␈αin␈α
general,␈αnot␈α
just␈αthe␈αencoding␈α
of
␈↓ α←␈↓strategy␈α
information.␈α
 We␈α
begin␈α
by␈α
examining␈α
in␈α
some␈α
detail␈α∞the␈α
difference
␈↓ α←␈↓between various approaches to referencing knowledge sources.

␈↓"β␈↓ α←␈↓␈↓α7-5-1    Reference by name vs. reference by description␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Since␈α⊂strategies␈α⊃(in␈α⊂any␈α⊂form)␈α⊃are␈α⊂used␈α⊂to␈α⊃direct␈α⊂the␈α⊃invocation␈α⊂of
␈↓ α←␈↓object-level␈α∀KSs,␈α∀they␈α∀must␈α∀have␈α∀some␈α∀way␈α∀of␈α∀referencing␈α∀them.␈α∪ Two
␈↓ α←␈↓fundamentally␈α∂different␈α∂approaches␈α∂have␈α∞typically␈α∂been␈α∂employed;␈α∂we␈α∞term
␈↓ α←␈↓them␈α␈↓↓reference␈αby␈α
name␈↓␈αand␈α␈↓↓reference␈α
by␈αdescription␈↓.␈α The␈α
former␈αlists␈αall␈α
KSs
␈↓ α←␈↓explicitly,␈α!while␈α"the␈α!latter␈α!offers␈α"a␈α!predicate␈α"indicating␈α!required
␈↓ α←␈↓characteristics.␈α∀ The␈α∀␈↓	THUSE␈↓␈α∀construct␈α∀of␈α∀␈↓¬PLANNER␈↓␈α∀uses␈α∀reference␈α∃by␈α∀name,
␈↓ α←␈↓allowing␈αthe␈αprogrammer␈αto␈αname␈αone␈αor␈αmore␈αtheorems␈αthat␈αare␈αlikely␈αto␈αbe
␈↓ α←␈↓especially useful.  The effect of a statement like

␈↓"β␈↓ α←␈↓	␈↓ β?(THGOAL (WIN POKER HAND) (THUSE BLUFF DRAW3 CHEAT))

␈↓ α←␈↓is␈αto␈αspecify␈αthe␈α
order␈αin␈αwhich␈αsome␈αof␈α
the␈αplausibly␈αuseful␈αtheorems␈αwill␈α
be
␈↓ α←␈↓applied.  The ␈↓	GOALCLASS␈↓ statement in ␈↓¬QA4␈↓ is quite similar.
␈↓"β␈↓ α←␈↓␈↓ β?Meta-rules,␈α⊂on␈α⊂the␈α⊂other␈α⊂hand,␈α⊂effect␈α⊂reference␈α⊂by␈α⊃description.␈α⊂ As
␈↓ α←␈↓previous␈α∂examples␈α∞have␈α∂illustrated,␈α∞they␈α∂make␈α∞conclusions␈α∂about␈α∞a␈α∂class␈α∞of
␈↓ α←␈↓rules␈α∂that␈α∂is␈α∂specified␈α∂by␈α∂describing␈α∂relevant␈α∂characteristics.␈α∂ ␈↓	METARULE002␈↓,
␈↓ α←␈↓for␈α∩instance,␈α∩refers␈α∩to␈α∩␈↓↓rules␈α∩which␈α∩mention␈α∩pseudomonas␈α∩in␈α∩their␈α⊃premise␈↓.
␈↓ α←␈↓␈↓¬PLANNER␈↓'s␈α∩theorem␈α⊃base␈α∩filter␈α∩(␈↓	THTBF␈↓)␈α⊃construct␈α∩is␈α⊃another␈α∩example␈α∩of␈α⊃this
␈↓ α←␈↓general␈α
sort␈αof␈α
capability: ␈α
It␈αallows␈α
the␈α
specification␈αof␈α
an␈αarbitrary␈α
predicate
␈↓ α←␈↓to filter the selection of theorems to be applied to a goal.
␈↓"β␈↓ α←␈↓␈↓ β?There␈αare␈α
numerous␈αadvantages␈αto␈α
reference␈αby␈αdescription,␈α
primarily
␈↓ α←␈↓in␈α∪terms␈α∪of␈α∪the␈α∪ability␈α∩of␈α∪the␈α∪knowledge␈α∪base␈α∪to␈α∪accommodate␈α∩changes.
␈↓ α←␈↓These␈α∃are␈α∃explored␈α∃later␈α∃in␈α∃Section␈α∃7-5-3.␈α∃ Here␈α∃we␈α∃examine␈α∃a␈α∀more
␈↓ α←␈↓detailed␈α≠point: ␈α≤the␈α≠implications␈α≠involved␈α≤in␈α≠two␈α≠different␈α≤ways␈α≠of
␈↓ α←␈↓accomplishing reference by description.
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    219␈↓

␈↓"β␈↓ α←␈↓␈↓α7-5-2    External descriptors vs. content reference␈↓
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∀way␈α∀to␈α∀accomplish␈α∀reference␈α∪by␈α∀description␈α∀is␈α∀via␈α∀a␈α∀set␈α∪of
␈↓ α←␈↓external␈α
descriptors.␈α
 That␈α
is,␈α
a␈αnumber␈α
of␈α
different␈α
characteristics␈α
could␈αbe
␈↓ α←␈↓chosen␈αand␈αeach␈αKS␈α
described␈αin␈αterms␈αof␈α
them.␈α For␈αan␈αordinary␈α
procedure,
␈↓ α←␈↓for␈α_instance,␈α_the␈α↔descriptor␈α_set␈α_might␈α↔include␈α_elements␈α_describing␈α↔the
␈↓ α←␈↓procedure's␈α∃main␈α⊗effect,␈α∃any␈α⊗side␈α∃effects,␈α∃preconditions␈α⊗for␈α∃its␈α⊗use,␈α∃etc.
␈↓ α←␈↓Strategies would then describe the relevant class of procedures in these terms.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂second␈α∞implementation␈α∂is␈α∞by␈α∂direct␈α∞examination␈α∂of␈α∂KS␈α∞content
␈↓ α←␈↓and␈α
is␈α
illustrated␈α∞by␈α
meta-rules.␈α
 For␈α∞instance,␈α
when␈α
␈↓	METARULE002␈↓␈α∞refers␈α
to
␈↓ α←␈↓␈↓↓rules␈α→that␈α_mention␈α→pseudomonas␈α→in␈α_their␈α→premise␈↓,␈α_the␈α→relevant␈α→set␈α_is
␈↓ α←␈↓determined␈α∂by␈α∞retrieving␈α∂the␈α∞premise␈α∂of␈α∞each␈α∂of␈α∞the␈α∂rules␈α∞in␈α∂question␈α∞and
␈↓ α←␈↓examining␈α⊂it␈α⊂directly␈α⊂to␈α⊂see␈α⊂if␈α∂it␈α⊂contains␈α⊂the␈α⊂desired␈α⊂item.␈↓
14␈↓␈α⊂That␈α⊂is,␈α∂the
␈↓ α←␈↓meta-rules␈α
examine␈α
the␈α
code␈α
of␈α∞the␈α
object-level␈α
rules␈α
to␈α
detect␈α∞the␈α
presence
␈↓ α←␈↓of any relevant characteristic.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α
general␈αnotion␈α
of␈α
allowing␈αone␈α
part␈α
of␈αthe␈α
system␈α
to␈αexamine␈α
the
␈↓ α←␈↓rules␈α
(code)␈α
executed␈α
by␈αother␈α
parts␈α
is␈α
based␈αon␈α
three␈α
main␈α
ideas.␈α First,␈α
there
␈↓ α←␈↓is␈α
the␈αconcept␈α
of␈αthe␈α
unity␈α
of␈αcode␈α
and␈αdata␈α
structure,␈αfirst␈α
suggested␈α
in␈αthe
␈↓ α←␈↓notion␈αof␈αa␈αstored␈αprogram␈α
computer␈αand␈αlater␈αmade␈αconvenient␈αby␈α
the␈α␈↓¬LISP␈↓
␈↓ α←␈↓language.␈α Second,␈αthe␈α
rules␈αmust␈αbe␈α
stored␈αin␈αa␈α
comprehensible␈αform.␈αIn␈α
this
␈↓ α←␈↓case␈α∩that␈α∩means␈α⊃interpreted␈α∩␈↓¬LISP␈↓␈α∩code,␈α⊃written␈α∩in␈α∩a␈α⊃stylized␈α∩form;␈α∩but␈α⊃in
␈↓ α←␈↓general␈α∞any␈α∞relatively␈α∞high-level␈α∂language␈α∞will␈α∞do.␈α∞ Finally,␈α∞the␈α∂syntax␈α∞and
␈↓ α←␈↓some␈α
of␈α
the␈α
semantics␈α
of␈αthat␈α
high-level␈α
language␈α
must␈α
be␈αrepresented␈α
within
␈↓ α←␈↓the␈α⊂system,␈α⊂to␈α⊂be␈α∂used␈α⊂as␈α⊂a␈α⊂guide␈α⊂in␈α∂examining␈α⊂the␈α⊂code.␈α⊂ In␈α⊂the␈α∂current
␈↓ α←␈↓example,␈α∃the␈α∃syntax␈α⊗is␈α∃represented␈α∃by␈α⊗the␈α∃template␈α∃for␈α⊗each␈α∃predicate
␈↓ α←␈↓function␈α∂(allowing␈α∂each␈α∂function␈α∂to␈α∞describe␈α∂its␈α∂own␈α∂calls)␈α∂and␈α∂the␈α∞stylized
␈↓ α←␈↓form␈α⊂of␈α∂rules.␈α⊂ Semantics␈α⊂are␈α∂supplied␈α⊂in␈α⊂part␈α∂by␈α⊂internal␈α⊂data␈α∂structures,
␈↓ α←␈↓which indicate, for example, that ␈↓	SAMEBUG␈↓ is an attribute.

␈↓"β␈↓ α←␈↓␈↓α7-5-3    Implications of content reference as an invocation mechanism␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈α
are␈α
a␈α
number␈α∞of␈α
interesting␈α
implications␈α
that␈α
follow␈α∞from␈α
the
␈↓ α←␈↓use␈αof␈αcontent-reference␈αas␈αan␈αinvocation␈αmechanism.␈α The␈α
technique␈αmakes
␈↓ α←␈↓possible␈α∞a␈α∞system␈α∞in␈α∞which␈α∂invocation␈α∞has␈α∞a␈α∞greater␈α∞degree␈α∞of␈α∂␈↓↓validity␈↓␈α∞and
␈↓ α←␈↓␈↓↓expressiveness␈↓␈α∂(defined␈α⊂below),␈α∂a␈α∂system␈α⊂that␈α∂makes␈α∂it␈α⊂easier␈α∂for␈α∂a␈α⊂user␈α∂to
␈↓ α←␈↓define␈α∂his␈α⊂own␈α∂␈↓↓generalized␈α∂invocation␈α⊂criteria␈↓,␈α∂and␈α∂a␈α⊂system␈α∂with␈α⊂a␈α∂higher
␈↓ α←␈↓degree␈αof␈α␈↓↓flexibility␈↓␈αin␈αresponding␈α
to␈αchanges.␈α We␈αconsider␈αeach␈αof␈α
these␈αin
␈↓ α←␈↓turn,␈α∂then␈α∂review␈α∂the␈α∂limitations␈α∞of␈α∂our␈α∂approach␈α∂and␈α∂the␈α∂difficulties␈α∞that
␈↓ α←␈↓remain before the benefits noted can be fully realized.

␈↓"β␈↓ α←␈↓␈↓αValidity and expressiveness␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Effecting␈αreference␈αby␈αdescription␈αthrough␈αdirect␈αexamination␈αof␈αrule
␈↓ α←␈↓content␈α∂can␈α∂be␈α∂seen␈α∂as␈α∂another␈α∂step␈α∂in␈α∂the␈α∂development␈α∂of␈α⊂KS␈α∂invocation
␈↓ α←␈↓mechanisms.␈α To␈αclarify␈αthis,␈α
consider␈αbreaking␈αdown␈αthe␈α
invocation␈αprocess
␈↓ α←␈↓into␈αtwo␈α
phases:␈αdetermining␈α
the␈α␈↓↓relevance␈↓␈α
and␈αdetermining␈α
the␈α␈↓↓applicability␈↓

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[14] This is accomplished via the template mechanism described in chapter 2.
␈↓ α←␈↓␈↓220    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓of␈α⊃a␈α⊃KS.␈α⊃The␈α⊃first␈α⊃involves␈α⊃the␈α⊃retrieval␈α⊃of␈α⊃one␈α⊃or␈α⊃more␈α⊃KSs␈α∩that␈α⊃may
␈↓ α←␈↓plausibly␈α→be␈α→useful;␈α→the␈α→second␈α→concerns␈α→the␈α→actual␈α→determination␈α_of
␈↓ α←␈↓applicability.␈α⊂ In␈α⊂the␈α⊃␈↓¬STRIPS␈↓␈α⊂system␈α⊂[Fikes71],␈α⊃for␈α⊂instance,␈α⊂relevance␈α⊃of␈α⊂an
␈↓ α←␈↓operator␈α∞is␈α∞determined␈α∞by␈α∞the␈α∞contents␈α∂of␈α∞the␈α∞add␈α∞and␈α∞delete␈α∞list,␈α∂while␈α∞its
␈↓ α←␈↓applicability is indicated by its preconditions.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αdetermination␈αof␈αrelevance␈αfor␈αa␈αKS␈αwill␈αbe␈αbased␈αon␈αsome␈αlink
␈↓ α←␈↓to␈αit,␈αa␈α␈↓↓handle␈↓␈αby␈αwhich␈αit␈α
can␈αbe␈αreferenced␈αand␈αretrieved␈α(e.g.,␈αin␈α␈↓¬STRIPS␈↓,␈α
the
␈↓ α←␈↓add␈α∂list).␈↓
15␈↓␈α∞We␈α∂will␈α∞be␈α∂concerned␈α∞here␈α∂primarily␈α∞with␈α∂the␈α∞various␈α∂types␈α∞of
␈↓ α←␈↓handles␈α⊂which␈α⊂have␈α⊂been␈α⊂used␈α⊂and,␈α⊂in␈α⊂particular,␈α⊂with␈α⊂their␈α⊂␈↓↓validity␈↓␈α∂and
␈↓ α←␈↓␈↓↓expressiveness␈↓␈α∀(defined␈α∀below).␈α∀ Table␈α∀7-1␈α∀sums␈α∀up␈α∀the␈α∃discussion␈α∀that
␈↓ α←␈↓follows.
␈↓"β␈↓ α←␈↓␈↓ β?Consider,␈α∪first,␈α∪the␈α∀types␈α∪of␈α∪handles␈α∪that␈α∀have␈α∪been␈α∪used.␈α∀ In␈α∪a
␈↓ α←␈↓historical␈αperspective,␈αthe␈αconcept␈αof␈αa␈αsubroutine␈α(procedure)␈αcan␈αbe␈αviewed
␈↓ α←␈↓as␈α∩the␈α∩introduction␈α⊃of␈α∩the␈α∩notion␈α⊃of␈α∩a␈α∩distinct,␈α⊃nontrivial␈α∩KS.␈α∩The␈α⊃only
␈↓ α←␈↓handle on it was its name, assigned by the programmer.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αfirst␈αmajor␈αdeparture␈αfrom␈αthis␈αcame␈αin␈α␈↓¬GPS␈↓␈α[Newell72],␈αand␈α␈↓¬GPS␈↓-
␈↓ α←␈↓like␈αsystems␈αsuch␈α
as␈α␈↓¬STRIPS␈↓.␈α In␈αthe␈α
latter,␈αfor␈αinstance,␈α
the␈αhandle␈αon␈αeach␈α
KS
␈↓ α←␈↓is␈αprovided␈α
by␈αthe␈α
contents␈αof␈αits␈α
add␈αand␈α
delete␈αlists.␈αNote␈α
that␈αpart␈α
of␈αthe
␈↓ α←␈↓␈↓↓definition␈↓␈αof␈α
the␈αKS␈αitself␈α
(the␈αadd␈αlist)␈α
provides␈αthe␈αhandle␈α
and␈αthe␈αname␈α
of
␈↓ α←␈↓the KS has become inconsequential.
␈↓"β␈↓ α←␈↓␈↓ β?Production␈α
rules␈α
are␈α
similar,␈α
since␈α
they␈α
are␈α
retrieved␈α
on␈α
the␈α
basis␈αof
␈↓ α←␈↓symbols␈α∂appearing␈α∞in␈α∂either␈α∂their␈α∞left-␈α∂or␈α∞right-hand␈α∂sides,␈α∂symbols␈α∞which
␈↓ α←␈↓are part of the definition of the KS itself.
␈↓"β␈↓ α←␈↓␈↓ β?With␈α∃the␈α∀advent␈α∃of␈α∀␈↓¬PLANNER␈↓-like␈α∃languages␈α∀(␈↓¬PLANNER␈↓,␈α∃␈↓¬QA4␈↓,␈α∃etc.),␈α∀the
␈↓ α←␈↓important␈α∩concepts␈α∪of␈α∩pattern-directed␈α∪and␈α∩goal-directed␈α∪invocation␈α∩were
␈↓ α←␈↓firmly␈α
established␈α
as␈αstandard␈α
programming␈α
language␈α
tools.␈α These␈α
languages
␈↓ α←␈↓provided␈αlinks␈αto␈αKSs␈αvia␈αpatterns␈αthat␈αwere␈αused␈αto␈αdescribe␈αeither␈αthe␈αgoal
␈↓ α←␈↓which␈α
the␈α
KS␈αcould␈α
achieve␈α
(consequent␈αtheorems)␈α
or␈α
the␈αevent␈α
to␈α
which␈αit
␈↓ α←␈↓was␈αrelevant␈α
(antecedent␈αand␈α
erasing␈αtheorems).␈α
 Once␈αagain,␈α
the␈αKS␈αname␈α
is
␈↓ α←␈↓irrelevant.
␈↓"β␈↓ α←␈↓␈↓ β?Consider␈α∩now␈α∪the␈α∩expressiveness␈α∪and␈α∩the␈α∪validity␈α∩of␈α∪the␈α∩handles
␈↓ α←␈↓provided␈α
by␈α
current␈α
programming␈α
techniques␈α
(see␈α
Fig.␈α
7-4)␈α
and␈α∞the␈α
impact
␈↓ α←␈↓of␈α⊃using␈α∩each␈α⊃of␈α∩them.␈α⊃ We␈α⊃define␈α∩␈↓↓expressiveness␈↓␈α⊃as␈α∩the␈α⊃richness␈α∩of␈α⊃the
␈↓ α←␈↓language␈αthat␈α
can␈αbe␈αused␈α
in␈αa␈αhandle.␈α
 One␈αcrude␈αmeasure␈α
is: ␈αDoes␈αit␈α
make
␈↓ α←␈↓any␈α
difference␈α
to␈α
program␈α
performance␈α
if␈α
every␈α
occurrence␈α
of␈α
a␈α∞KS␈α
handle
␈↓ α←␈↓in␈αthe␈αprogram␈αtext␈αis␈αuniformly␈αreplaced␈αby␈αsome␈αarbitrary␈αstring?␈α That␈αis,
␈↓ α←␈↓is the handle merely a token or does its structure convey information?
␈↓"β␈↓ α←␈↓␈↓ β?The␈α↔␈↓↓validity␈↓␈α↔of␈α⊗a␈α↔handle␈α↔is␈α↔determined␈α⊗by␈α↔the␈α↔nature␈α↔of␈α⊗the

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[15]␈α(We␈αdistinguish␈αbetween␈αthe␈α␈↓↓handle␈↓␈αand␈α␈↓↓body␈↓␈αof␈αa␈αKS: ␈αThe␈αbody␈αis␈αthe
␈↓ α←␈↓part␈α∞actually␈α
executed␈α∞[or␈α∞interpreted],␈α
while␈α∞the␈α∞handle␈α
is␈α∞anything␈α∞that␈α
is
␈↓ α←␈↓used␈αas␈αa␈αway␈αof␈αaccessing␈αthe␈αbody.␈αThus,␈αfor␈αa␈αsubroutine,␈αthe␈αname␈αis␈αthe
␈↓ α←␈↓handle␈α
and␈α∞its␈α
code␈α∞is␈α
the␈α∞body;␈α
while␈α
for␈α∞a␈α
␈↓¬PLANNER␈↓␈α∞theorem,␈α
the␈α∞pattern␈α
is
␈↓ α←␈↓the␈α⊂handle␈α⊂and␈α⊂the␈α⊂theorem␈α⊂code␈α⊃is␈α⊂the␈α⊂body.␈α⊂ As␈α⊂will␈α⊂become␈α⊃clear,␈α⊂the
␈↓ α←␈↓handle and the body are not necessarily distinct.)
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    221␈↓

␈↓"β␈↓ α←␈↓relationship␈α
between␈αthe␈α
handle␈α
and␈αthe␈α
KS␈α
body.␈α A␈α
crude␈α
measure␈αmight
␈↓ α←␈↓be␈α∞the␈α∞effect␈α∞of␈α∞editing␈α∞the␈α∞KS: ␈α∞Given␈α∞some␈α∞KS␈α∞invoked␈α∞once␈α∂during␈α∞the
␈↓ α←␈↓course␈α
of␈α
program␈α
execution,␈α
what␈α
effect␈α
will␈α
editing␈α
the␈α
KS␈α
body␈α∞have␈α
on
␈↓ α←␈↓that␈α
invocation?␈α
 That␈α
is,␈α∞will␈α
editing␈α
just␈α
the␈α∞KS␈α
body␈α
have␈α
any␈α∞effect␈α
on
␈↓ α←␈↓whether or not it is ever invoked?
␈↓"β␈↓ α←␈↓␈↓ β?For␈α∩example,␈α∩the␈α∩␈↓↓name␈↓␈α∩of␈α∩a␈α∩subroutine␈α∩is␈α∩purely␈α∩a␈α∩token␈α∩and␈α⊃is
␈↓ α←␈↓devoid␈α∞of␈α∞semantics␈α∞(except␈α∞of␈α∞course␈α∞in␈α∞the␈α∞programmer's␈α∞mind).␈α∞There␈α∞is
␈↓ α←␈↓only␈α∂one␈α∞way␈α∂to␈α∞express␈α∂which␈α∂subroutine␈α∞you␈α∂want␈α∞(by␈α∂naming␈α∂it);␈α∞hence
␈↓ α←␈↓this␈α_technique␈α_has␈α_minimal␈α_expressiveness.␈α_ Since␈α_there␈α_is␈α_no␈α↔formal
␈↓ α←␈↓connection␈αbetween␈αthe␈αname␈αand␈α
contents,␈αarbitrary␈αchanges␈αcan␈αbe␈αmade␈α
to
␈↓ α←␈↓the␈α∂body␈α∂without␈α∂affecting␈α∂whether␈α∂or␈α∞not␈α∂it␈α∂is␈α∂invoked.␈α∂ In␈α∂this␈α∂case,␈α∞the
␈↓ α←␈↓handle␈α∩has␈α⊃minimal␈α∩validity.␈α∩ As␈α⊃a␈α∩result,␈α⊃the␈α∩programmer␈α∩himself␈α⊃must
␈↓ α←␈↓know␈αexactly␈αwhat␈α
each␈αsubroutine␈αcontains␈αand␈α
select,␈αby␈αname,␈α
the␈αproper
␈↓ α←␈↓one at the proper point in his program.
␈↓"β␈↓ α←␈↓␈↓ β?With␈α_the␈α↔advent␈α_of␈α↔pattern-directed␈α_invocation,␈α_the␈α↔subroutine
␈↓ α←␈↓acquired␈αa␈α
␈↓↓pattern␈↓␈αwhich␈α
is␈αused␈αas␈α
the␈αhandle.␈α
There␈αis␈α
a␈αlimited␈αbut␈α
useful
␈↓ α←␈↓sort␈αof␈α
expressiveness␈αhere,␈α
in␈αa␈αsyntax␈α
that␈αtypically␈α
permits␈αa␈α
broad␈αrange
␈↓ α←␈↓of␈α∂patterns.␈α∂ The␈α⊂structure␈α∂of␈α∂the␈α⊂pattern␈α∂conveys␈α∂information␈α⊂and,␈α∂hence,
␈↓ α←␈↓can␈α∪not␈α∪in␈α∩general␈α∪be␈α∪replaced␈α∪by␈α∩an␈α∪arbitrary␈α∪string␈α∪without␈α∩affecting
␈↓ α←␈↓program␈αperformance.␈α Consider,␈αhowever,␈αthe␈αvalidity␈αof␈αthe␈αlink.␈αNote␈αthat
␈↓ α←␈↓it␈αis␈α
the␈αprogrammer␈αwho␈α
writes␈αboth␈αthe␈α
body␈αand␈αthe␈α
pattern␈αfor␈α
the␈αKS,
␈↓ α←␈↓hence␈α∩there␈α∩is␈α∩no␈α∪guaranteed␈α∩correspondence.␈α∩ Arbitrary␈α∩changes␈α∪can␈α∩be
␈↓ α←␈↓made␈α∪to␈α∀the␈α∪body␈α∀of␈α∪a␈α∀␈↓¬PLANNER␈↓␈α∪theorem,␈α∀for␈α∪instance,␈α∀without␈α∪changing
␈↓ α←␈↓whether or not it is invoked.
␈↓"β␈↓ α←␈↓␈↓ β?Goal-directed␈α!invocation␈α was␈α!another␈α step␈α!toward␈α increased
␈↓ α←␈↓expressiveness␈α
as␈α
the␈α
pattern␈α
acquired␈α
an␈α
interpretation,␈α
which␈α
added␈α
further
␈↓ α←␈↓richness␈α∞to␈α∞the␈α∞language.␈α∞ In␈α∞␈↓¬PLANNER␈↓,␈α∞for␈α∞instance,␈α∞there␈α∞are␈α∞three␈α∞classes␈α∞of
␈↓ α←␈↓patterns␈α⊗(goal,␈α∃assertion␈α⊗event,␈α∃and␈α⊗erasing␈α∃event),␈α⊗while␈α∃␈↓¬QA4␈↓␈α⊗offers␈α∃an
␈↓ α←␈↓extensive␈α
set␈α
of␈α
categories␈α
in␈αa␈α
general␈α
demon␈α
mechanism.␈α
 But␈α
the␈αvalidity
␈↓ α←␈↓remains␈α⊃the␈α⊃same␈α⊂(i.e.,␈α⊃minimal).␈α⊃ Indeed,␈α⊃there␈α⊂is␈α⊃the␈α⊃potential␈α⊃for␈α⊂much
␈↓ α←␈↓deviousness␈α∀in␈α∀allowing␈α∀the␈α∀programmer␈α∀to␈α∀cause␈α∀the␈α∀invocation␈α∀of␈α∪an
␈↓ α←␈↓arbitrary␈α∞body␈α∞of␈α∞code␈α
based␈α∞on␈α∞a␈α∞pattern␈α
which␈α∞may␈α∞or␈α∞may␈α∞not␈α
describe
␈↓ α←␈↓what the theorem actually achieves:␈↓
16␈↓
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[16]␈αAnd␈α
herein␈αlies␈α
a␈αseductive␈α
trap.␈α The␈αKS␈α
handle␈αis␈α
no␈αlonger␈α
a␈αtoken
␈↓ α←␈↓but␈α_conveys␈α→information.␈α_ When␈α→that␈α_information␈α→is␈α_advertised␈α→as␈α_a
␈↓ α←␈↓``purpose,''␈α∂it␈α∞is␈α∂easy␈α∞to␈α∂start␈α∂believing␈α∞that␈α∂every␈α∞KS␈α∂is␈α∞sure␈α∂to␈α∂achieve␈α∞its
␈↓ α←␈↓advertised␈α∞``purpose.'' ␈α∞Consider␈α
the␈α∞psychological␈α∞difference␈α∞between␈α
asking
␈↓ α←␈↓for␈α
``any␈α∞KS␈α
whose␈α∞pattern␈α
matches␈α
the␈α∞following''␈α
and␈α∞asking␈α
for␈α∞``any␈α
KS
␈↓ α←␈↓that␈α∂achieves␈α∂the␈α∞following␈α∂goal.'' ␈α∂Note␈α∞that␈α∂the␈α∂computational␈α∞mechanism
␈↓ α←␈↓for␈αboth␈αis␈αidentical␈αand␈αthe␈α
difference␈α(the␈αinterpretation␈αof␈αthe␈αpattern␈αas␈α
a
␈↓ α←␈↓goal)␈α⊃exists␈α⊃purely␈α⊃in␈α⊃the␈α⊂programmer's␈α⊃mind.␈α⊃ The␈α⊃first␈α⊃asks␈α⊃the␈α⊂proper
␈↓ α←␈↓question,␈α∂in␈α∂that␈α∞it␈α∂assumes␈α∂nothing␈α∞more␈α∂than␈α∂it␈α∞can␈α∂deliver;␈α∂the␈α∂latter␈α∞is
␈↓ α←␈↓fraught␈αwith␈αeasily␈αmisinterpreted␈αconnotations.␈α It␈αis␈αeasy␈αto␈αforget␈αthat␈αit␈αis
␈↓ α←␈↓the␈αprogrammer's␈α
responsibility␈αto␈αmake␈α
sure␈αeach␈α
KS␈αachieves␈αits␈α
advertised
␈↓ α←␈↓effect.
␈↓ α←␈↓␈↓222    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓␈↓ β'The␈α
subroutines␈αare␈α
not␈αreferenced␈α
by␈αtheir␈α
names.␈α
 Instead␈αthey
␈↓ α←␈↓␈↓ β'are␈αcalled␈α
because␈αthey␈αaccept␈α
arguments␈αwith␈αa␈α
certain␈αstructure,
␈↓ α←␈↓␈↓ β'and␈α
because␈α
␈↓↓the␈α
programmer␈α
claimed␈↓␈αthat␈α
they␈α
will␈α
solve␈α
goals␈αof␈α
a
␈↓ α←␈↓␈↓ β'certain class.
␈↓"β␈↓ α←␈↓␈↓ β'␈↓ ¬,from the ␈↓¬QA4␈↓ manual [Rulifson72], (emphasis added) 

␈↓"β␈↓ α←␈↓␈↓ β?As␈α⊃the␈α∩final␈α⊃step␈α∩in␈α⊃this␈α∩dimension␈α⊃we␈α⊃have␈α∩the␈α⊃use␈α∩of␈α⊃a␈α∩set␈α⊃of
␈↓ α←␈↓descriptors,␈α∂the␈α∂``external␈α∂descriptor''␈α∂approach␈α∂described␈α∂earlier.␈α∂ Note␈α∞that
␈↓ α←␈↓this␈α
is␈α
a␈α
generalization␈α
of␈α
goal-directed␈α
invocation,␈α
since␈α
the␈α
``purpose''␈α
of␈α
a
␈↓ α←␈↓KS␈α∞is␈α∞only␈α∞one␈α∞description␈α∞of␈α∞it.␈α∞ Any␈α∞number␈α∞of␈α∞other␈α∞facts␈α∞about␈α∞it␈α∞may
␈↓ α←␈↓prove␈α
relevant␈α
and␈α
could,␈α
equally␈α
well,␈α
be␈α
supplied.␈α
 A␈α
suitably␈α
large␈α∞set␈α
of
␈↓ α←␈↓such␈α
descriptors␈α
would␈αmake␈α
possible␈α
a␈α
useful␈αlanguage␈α
for␈α
referencing␈αa␈α
KS.
␈↓ α←␈↓As␈α␈↓↓external␈↓␈αdescriptors,␈αhowever,␈αthey␈α
have␈αno␈αformal␈αrelationship␈αto␈αthe␈α
KS
␈↓ α←␈↓body,␈αand␈αtheir␈α
validity␈αremains␈αminimal--arbitrary␈α
changes␈αto␈αthe␈αbody␈α
will
␈↓ α←␈↓not automatically modify the descriptors or the invocation pattern of the KS.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
link␈αin␈α
␈↓¬GPS␈↓␈αand␈α
␈↓¬STRIPS␈↓␈αis,␈α
as␈α
noted,␈αbased␈α
on␈αpart␈α
of␈αthe␈α
definition
␈↓ α←␈↓of␈αthe␈αKS␈αitself␈αand␈αthus␈αhas␈αa␈αmuch␈αstronger␈αdegree␈αof␈αvalidity.␈α However,
␈↓ α←␈↓all␈αof␈α
the␈αretrieval␈α
mechanism␈αis␈αcontained␈α
in␈αthe␈α
means-ends␈αanalysis␈αthat␈α
is
␈↓ α←␈↓embedded␈αin␈αthe␈αsystem␈αinterpreter␈αand␈αthe␈αprogrammer␈αhas␈αno␈αcontrol␈αover
␈↓ α←␈↓which␈α
KS␈α∞is␈α
invoked.␈α
Thus,␈α∞while␈α
the␈α
handle␈α∞on␈α
the␈α
KS␈α∞is␈α
valid,␈α∞the␈α
user
␈↓ α←␈↓has␈α∞no␈α∞means␈α∞of␈α∞expressing␈α∂his␈α∞preferences␈α∞for␈α∞which␈α∞should␈α∂be␈α∞retrieved.
␈↓ α←␈↓Production␈α∂rules␈α∞are␈α∂similar: ␈α∞The␈α∂link␈α∞is␈α∂valid␈α∞because␈α∂it␈α∞is␈α∂based␈α∂on␈α∞the
␈↓ α←␈↓content␈α⊗of␈α⊗the␈α⊗KS,␈α⊗but␈α⊗there␈α⊗is␈α⊗the␈α⊗same␈α⊗sort␈α⊗of␈α⊗single,␈α∃``hardwired''
␈↓ α←␈↓mechanism␈α∪that␈α∩effects␈α∪the␈α∪retrieval,␈α∩leaving␈α∪the␈α∩user␈α∪no␈α∪opportunity␈α∩to
␈↓ α←␈↓express a preference.␈↓
17␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Note␈α
that␈αin␈α
both␈α
of␈αthese␈α
cases␈α
(␈↓¬GPS␈↓␈αand␈α
production␈α
rules),␈αediting␈α
the
␈↓ α←␈↓body␈α→of␈α→the␈α→KS␈α→alone␈α→will␈α→produce␈α→changes␈α→in␈α→invocation␈α_patterns,
␈↓ α←␈↓demonstrating the higher level of validity of the handle.
␈↓"β␈↓ α←␈↓␈↓ β?Traditional␈α∀invocation␈α∀mechanisms␈α∪thus␈α∀offer␈α∀varying␈α∀degrees␈α∪of
␈↓ α←␈↓expressiveness␈α∞and␈α∂validity.␈α∞ Is␈α∞it␈α∂possible␈α∞to␈α∞obtain␈α∂both␈α∞of␈α∞these␈α∂at␈α∞once?
␈↓ α←␈↓The␈α∂current␈α∂implementation␈α∂of␈α∂meta-rules␈α∂uses␈α∂a␈α∂technique␈α∂that␈α∂takes␈α∞one
␈↓ α←␈↓step␈α
in␈α
this␈α
direction;␈α
we␈α
term␈α
it␈α
␈↓↓content-directed␈α
invocation␈↓␈α
to␈α
suggest␈α
its␈α
place
␈↓ α←␈↓in the ongoing development of KS invocation methods.␈↓
18␈↓

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[17]␈αIt␈αis␈αpossible␈αto␈αforce␈αspecified␈αinteractions␈αby␈αanticipating␈αthe␈αoperation
␈↓ α←␈↓of␈αthe␈αretrieval␈α
mechanism,␈αbut␈αthis␈α
is␈αoften␈αdifficult␈α
for␈αa␈αlarge␈α
system␈αand
␈↓ α←␈↓is,␈α⊂in␈α⊂any␈α⊂case,␈α⊂contrary␈α⊃to␈α⊂the␈α⊂``spirit''␈α⊂of␈α⊂the␈α⊂formalism.␈α⊃ See␈α⊂[Davis77a],
␈↓ α←␈↓especially Section 5, for a discussion of the issue.

␈↓"β␈↓ α←␈↓[18]␈αThere␈αhas␈αbeen␈αa␈αparallel␈αevolution␈αin␈αaccess␈αto␈αmemory␈αlocations,␈αfrom
␈↓ α←␈↓absolute␈α∩binary,␈α∩to␈α∩symbolic␈α∩addressing␈α∩in␈α∩assemblers,␈α∩to␈α∩relocatable␈α∩core
␈↓ α←␈↓images,␈α∀on␈α∃up␈α∀to␈α∀content␈α∃addressable␈α∀memories␈α∀(e.g.,␈α∃␈↓¬LEAP␈↓␈α∀[Feldman72]).
␈↓ α←␈↓Content-directed␈α
invocation␈α
can␈αthus␈α
be␈α
seen␈αas␈α
the␈α
procedural␈α
analogue␈αof
␈↓ α←␈↓content addressable memory.
␈↓"β␈↓ α←␈↓␈↓ β?An␈α∨alternative␈α∨to␈α content-directed␈α∨invocation␈α∨would␈α be␈α∨the
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    223␈↓

␈↓"β␈↓ α←␈↓␈↓ β?It␈α∂has␈α∂a␈α∂high␈α∂degree␈α∂of␈α∂validity␈α∂because,␈α∂like␈α∂production␈α⊂rules␈α∂and
␈↓ α←␈↓␈↓¬GPS␈↓-like␈α
systems,␈αit␈α
references␈α
the␈αKS␈α
code␈α
directly.␈α Recall␈α
that␈α
by␈α``validity''
␈↓ α←␈↓we␈αmean␈αthat␈αthe␈α
KS␈αretrieved␈αwill␈αin␈α
fact␈αachieve␈αthe␈αdesired␈α
effect.␈α That
␈↓ α←␈↓is,␈α␈↓	METARULE002␈↓,␈αfor␈αexample␈α(Fig.␈α7-1),␈αwill␈αretrieve␈αan␈αobject-level␈αrule␈αnot
␈↓ α←␈↓because␈αof␈α
the␈αrule␈α
name,␈αthe␈αrule␈α
number,␈αor␈α
other␈αexternal␈α
descriptor,␈αbut
␈↓ α←␈↓because␈α
examination␈α
of␈α
the␈α
code␈α
of␈α
that␈α
object-level␈α
rule␈α
reveals␈α
that␈α
the␈α
rule
␈↓ α←␈↓does␈α⊂in␈α⊂fact␈α⊂conclude␈α⊂about␈α⊂enterobacteriaceae.␈α⊂ The␈α⊂meta-rules␈α⊂cannot␈α⊂of
␈↓ α←␈↓course␈α∀assure␈α∀that␈α∀the␈α∃object-level␈α∀rule␈α∀makes␈α∀this␈α∀conclusion␈α∃with␈α∀the
␈↓ α←␈↓appropriate␈α⊂justification--it␈α⊂would␈α⊂be␈α⊂far␈α⊂more␈α⊂difficult␈α⊂for␈α⊂the␈α⊂system␈α⊂to
␈↓ α←␈↓determine␈αthat␈αa␈αKS␈αnot␈αonly␈αachieves␈αa␈αdesired␈αeffect␈αbut␈αdoes␈αso␈αbased␈αon
␈↓ α←␈↓reasoning␈α∂(or␈α∂actions)␈α∂appropriate␈α∂to␈α∂and␈α∂justifiable␈α∂in␈α∂the␈α∂domain.␈α∞ Thus
␈↓ α←␈↓whether␈α⊂the␈α⊂object-level␈α⊂rule␈α⊂deduces␈α⊂enterobacteriaceae␈α⊂for␈α⊂valid␈α⊂medical
␈↓ α←␈↓reasons␈αis␈αa␈α
different␈αissue␈αthat␈α
would␈αrequire␈αa␈α
great␈αdeal␈αmore␈αinference␈α
on
␈↓ α←␈↓the␈α⊃part␈α⊃of␈α⊃the␈α⊃system.␈α⊃ We␈α∩claim␈α⊃to␈α⊃have␈α⊃taken␈α⊃only␈α⊃a␈α⊃first␈α∩step␈α⊃(from
␈↓ α←␈↓external␈α
descriptors␈α
to␈α∞content-reference)␈α
but␈α
submit␈α
that␈α∞it␈α
can␈α
be␈α∞a␈α
useful
␈↓ α←␈↓advance, especially in a knowledge base undergoing frequent revision.
␈↓"β␈↓ α←␈↓␈↓ β?Content-directed␈αinvocation␈αis␈α
also␈αexpressive,␈αsince␈α
it␈αoffers␈αa␈αway␈α
of
␈↓ α←␈↓using␈αany␈αcomputable␈αpredicate␈αthe␈αuser␈αdefines␈α(e.g.,␈α␈↓	MENTIONS␈↓␈αin␈αFig.␈α7-1).
␈↓ α←␈↓The␈αlink␈αthus␈αinherits␈αits␈αexpressiveness␈αfrom␈αthe␈αprogramming␈αlanguage␈αin
␈↓ α←␈↓which the system is written.
␈↓"β␈↓ α←␈↓␈↓ β?Consider␈α∀once␈α∀again␈α∪the␈α∀historical␈α∀perspective.␈α∀ The␈α∪programmer
␈↓ α←␈↓using␈α⊂procedures␈α⊂effectively␈α∂says␈α⊂``give␈α⊂me␈α⊂that␈α∂KS␈α⊂next,''␈α⊂indicating␈α⊂it␈α∂by

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓formalization␈α∞of␈α∞the␈α∂link␈α∞between␈α∞the␈α∞external␈α∂descriptors␈α∞and␈α∞the␈α∂body␈α∞of
␈↓ α←␈↓the␈αKS.␈α There␈αare␈αat␈αleast␈αtwo␈αways␈αthis␈αmapping␈αmight␈αbe␈αprovided.␈α The
␈↓ α←␈↓first␈αis␈αa␈α``descriptor␈αverifier''␈αapproach.␈α This␈αwould␈αinvolve␈αsubmitting␈αto␈αa
␈↓ α←␈↓deductive␈α∞system␈α∞both␈α∞the␈α∞code␈α∞for␈α∞a␈α∞KS␈α∞and␈α∞some␈α∞descriptors␈α∞for␈α∞it,␈α∞then
␈↓ α←␈↓allowing␈α∩the␈α⊃deductive␈α∩system␈α⊃to␈α∩attempt␈α⊃to␈α∩prove␈α⊃the␈α∩correctness␈α∩of␈α⊃the
␈↓ α←␈↓descriptors␈α∃(much␈α⊗like␈α∃current␈α∃work␈α⊗in␈α∃program␈α⊗verification).␈α∃ Another
␈↓ α←␈↓scheme␈α∞would␈α∞be␈α∞to␈α∞use␈α∞a␈α∞``descriptor␈α∞generator,''␈α∞which,␈α∞provided␈α∂with␈α∞the
␈↓ α←␈↓KS␈α≥code␈α≥and␈α≥a␈α≥characterization␈α≥of␈α≥the␈α≥descriptors␈α≥desired,␈α≤would
␈↓ α←␈↓automatically␈α∞generate␈α∞them.␈α
 A␈α∞simple␈α∞version␈α∞of␈α
this␈α∞was␈α∞done␈α∞some␈α
time
␈↓ α←␈↓ago:␈α⊃GPS␈α⊃was␈α⊃capable␈α⊃of␈α⊂constructing␈α⊃its␈α⊃own␈α⊃table␈α⊃of␈α⊃connections␈α⊂when
␈↓ α←␈↓supplied␈α∩with␈α∩operators␈α⊃in␈α∩the␈α∩form␈α∩of␈α⊃rewrite␈α∩rules␈α∩(e.g.,␈α∩symbolic␈α⊃logic
␈↓ α←␈↓transformation␈α~rules)␈α~and␈α~a␈α≠set␈α~of␈α~routines␈α~for␈α≠defining␈α~differences
␈↓ α←␈↓([Newell59],␈α∂[Newell61]).␈α∂ It␈α∂matched␈α∂the␈α∂left-␈α∂and␈α∂right-hand␈α∂sides␈α⊂of␈α∂the
␈↓ α←␈↓rules␈α∂and␈α∞applied␈α∂each␈α∂difference␈α∞detector␈α∂to␈α∂the␈α∞result.␈α∂ This␈α∂was␈α∞feasible
␈↓ α←␈↓because␈α~the␈α≠KS␈α~``code''␈α≠had␈α~a␈α≠very␈α~simple␈α≠form,␈α~and␈α≠because␈α~the
␈↓ α←␈↓``characterization␈α⊃of␈α⊃descriptors''␈α⊂was␈α⊃a␈α⊃procedure␈α⊂for␈α⊃finding␈α⊃the␈α⊂relevant
␈↓ α←␈↓features.
␈↓"β␈↓ α←␈↓␈↓ β?More␈α⊂powerful␈α⊂versions␈α⊂of␈α∂either␈α⊂of␈α⊂these,␈α⊂however,␈α⊂would␈α∂require
␈↓ α←␈↓substantial␈αadvances␈αin␈αthe␈αstate␈αof␈αthe␈αart.␈α It␈αseems␈αmore␈αreasonable␈αfor␈α
the
␈↓ α←␈↓time␈α∂being␈α∂to␈α∂use␈α∂content-directed␈α∂invocation,␈α∂which␈α∂employs␈α∂a␈α∂procedural
␈↓ α←␈↓definition␈αof␈αthe␈αcharacteristic␈αdesired␈αand␈αallows␈αthe␈αprocedure␈αto␈αaccess␈αthe
␈↓ α←␈↓KS code, as the function ␈↓	MENTIONS␈↓ does.
␈↓ α←␈↓␈↓224    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓name.␈α In␈α␈↓¬GPS-STRIPS␈↓␈αand␈αtraditional␈αproduction␈αsystems,␈αthe␈αuser␈αhas␈αlittle␈αor
␈↓ α←␈↓no␈α∞control␈α∞over␈α∞which␈α
KS␈α∞is␈α∞invoked␈α∞next.␈α
 ␈↓¬PLANNER␈↓␈α∞made␈α∞it␈α∞possible␈α∞to␈α
say
␈↓ α←␈↓``give␈αme␈αany␈αKS␈αwhose␈αpattern␈αmatches␈αthis␈αone'';␈αin␈αusing␈αthat␈αpattern␈αas␈αa
␈↓ α←␈↓designator␈αof␈αa␈αgoal,␈αthe␈αrequest␈αbecomes␈α``give␈αme␈αany␈αKS␈αthat␈α(supposedly)
␈↓ α←␈↓achieves␈α∂the␈α∂goal␈α∂designated.''␈α∞Finally,␈α∂content-directed␈α∂invocation␈α∂makes␈α∞it
␈↓ α←␈↓possible␈α⊃to␈α⊂say␈α⊃``give␈α⊂me␈α⊃any␈α⊃KS␈α⊂that␈α⊃fits␈α⊂the␈α⊃following␈α⊃description.'' ␈α⊂By
␈↓ α←␈↓writing␈α
the␈α
proper␈α
sort␈α
of␈α
description,␈α
we␈α
can␈α
have␈α
invocation␈α
that␈α
is␈α
goal-
␈↓ α←␈↓directed,␈α
side-effect-directed,␈α
speed-directed--in␈αshort,␈α
directed␈α
by␈α
any␈αone␈α
or
␈↓ α←␈↓a combination of factors.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α⊃of␈α⊃the␈α⊃interesting␈α⊃contributions␈α⊃␈↓¬PLANNER␈↓␈α⊃made␈α⊃was␈α⊃to␈α⊃take␈α⊃the
␈↓ α←␈↓notion␈αof␈αgoal-directed␈αretrieval␈αout␈α
of␈αthe␈αinterpreter␈α(as␈αin␈α␈↓¬STRIPS␈↓)␈α
and␈αput
␈↓ α←␈↓it␈α
in␈α
the␈α
hands␈α
of␈α
the␈α
programmer.␈α
 In␈α
a␈α
similar␈α
vein,␈α
we␈α
have␈α∞offered␈α
the
␈↓ α←␈↓programmer␈α
the␈α
notion␈α
of␈α
retrieval␈α
itself: ␈α
The␈α
criteria␈α
for␈α
KS␈α
retrieval␈α
are
␈↓ α←␈↓no␈α∞longer␈α∞␈↓↓predefined␈↓␈α∞and␈α∞␈↓↓embedded␈α∞in␈α∞the␈α∞language␈α∞interpreter␈↓,␈α∞but␈α∞can␈α
be
␈↓ α←␈↓specified and changed (even dynamically) by the user himself.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α
can␈α
be␈α
pursued␈α
a␈α
step␈α
further.␈α
 The␈α
second␈α
clause␈α
of␈α
the␈α
premise
␈↓ α←␈↓of the meta-rule in Fig. 7-1

␈↓"β␈↓ α←␈↓	␈↓ β∂(THEREARE OBJRULES(MENTIONS CNTXT PREMISE SAMEBUG) SET1))

␈↓ α←␈↓is␈α⊃one␈α⊃very␈α⊃simple␈α∩example␈α⊃of␈α⊃content-directed␈α⊃invocation.␈α∩ It␈α⊃determines
␈↓ α←␈↓whether␈α→or␈α_not␈α→an␈α_object-level␈α→rule␈α_mentions␈α→the␈α→attribute␈α_␈↓	SAMEBUG␈↓.
␈↓ α←␈↓Consider␈α∞then␈α∞the␈α∞impact␈α∞of␈α∞second␈α∞order␈α∞meta-rules--they␈α∞can␈α∞be␈α∂used␈α∞to
␈↓ α←␈↓select␈αthe␈αcriteria␈αby␈αwhich␈αthe␈αKSs␈αwill␈αbe␈αcharacterized.␈α Thus␈αwe␈αnot␈αonly
␈↓ α←␈↓allow␈α∂the␈α∂user␈α∞to␈α∂specify␈α∂arbitrary␈α∞criteria␈α∂to␈α∂control␈α∞retrieval,␈α∂but␈α∂make␈α∞it
␈↓ α←␈↓possible␈αfor␈αhim␈α
to␈αencode␈αknowledge␈αthat␈α
decides␈αdynamically␈αwhich␈αare␈α
the
␈↓ α←␈↓most appropriate criteria to use.
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    225␈↓



␈↓ α←␈↓α␈↓ ∧2Table 7-1.    KS Invocation Mechanisms.    



␈↓"␈↓ α←␈↓∧αααααααααααααπααααααααααααααααπααααααααααααπααααααααααααααααααα
␈↓"␈↓ α←␈↓∧KS Type      ~ Type of Handle ~ Validity   ~ Expressiveness
␈↓"␈↓ α←␈↓∧αααααααααααααβααααααααααααααααβααααααααααααβααααααααααααααααααα
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧Subroutine   ~ Name           ~ Minimal    ~ Minimal
␈↓"␈↓ α←␈↓∧GPS - STRIPS ~ Add list       ~ Strong     ~ Very limited,
␈↓"␈↓ α←␈↓∧             ~                ~            ~ hardwired
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧Production   ~ Symbols in     ~ Strong     ~ Very limited,
␈↓"␈↓ α←␈↓∧rule         ~ rule           ~            ~ hardwired
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧PLANNER      ~ Pattern +      ~ Minimal    ~ 3 categories (goal
␈↓"␈↓ α←␈↓∧theorem      ~ recommendation ~            ~ assertion, erasing
␈↓"␈↓ α←␈↓∧             ~ list advice    ~            ~ and limited syntax
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧QA4          ~ Pattern +      ~ Minimal    ~ Syntax similar to
␈↓"␈↓ α←␈↓∧operator     ~ GOALCLASS      ~            ~ PLANNER, but wider
␈↓"␈↓ α←␈↓∧             ~ advice         ~            ~ set of categories
␈↓"␈↓ α←␈↓∧             ~                ~            ~ in a general demon
␈↓"␈↓ α←␈↓∧             ~                ~            ~ mechanism
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧             ~                ~            ~
␈↓"␈↓ α←␈↓∧Content-     ~ Arbitrary      ~ Explicit,  ~ Extensive,
␈↓"␈↓ α←␈↓∧directed     ~ predicate on   ~ formalized,~ extensible,
␈↓"␈↓ α←␈↓∧invocation   ~ KS contents    ~ testable   ~ softwired
␈↓"␈↓ α←␈↓∧ααααααααααααα∀αααααααααααααααα∀αααααααααααα∀ααααααααααααααααααα
␈↓ α←␈↓␈↓226    STRATEGIES␈↓ 
#7-5␈↓









␈↓"␈↓ α←␈↓∧       ↑
␈↓"␈↓ α←␈↓∧       ~                                Content-directed
␈↓"␈↓ α←␈↓∧       ~                                invocation
␈↓"␈↓ α←␈↓∧       ~   Set of external
␈↓"␈↓ α←␈↓∧       ~   descriptors
␈↓"␈↓ α←␈↓∧       ~
␈↓"␈↓ α←␈↓∧E      ~
␈↓"␈↓ α←␈↓∧X      ~
␈↓"␈↓ α←␈↓∧P      ~   Goal-directed
␈↓"␈↓ α←␈↓∧R      ~   invocation
␈↓"␈↓ α←␈↓∧E      ~
␈↓"␈↓ α←␈↓∧S      ~
␈↓"␈↓ α←␈↓∧S      ~
␈↓"␈↓ α←␈↓∧I      ~   Pattern-directed
␈↓"␈↓ α←␈↓∧V      ~   invocation
␈↓"␈↓ α←␈↓∧E      ~
␈↓"␈↓ α←␈↓∧N      ~
␈↓"␈↓ α←␈↓∧E      ~
␈↓"␈↓ α←␈↓∧S      ~   Subroutine                   GPS, STRIPS
␈↓"␈↓ α←␈↓∧S      ~                                production rules
␈↓"␈↓ α←␈↓∧       ~
␈↓"␈↓ α←␈↓∧       %ααααααααααααααααααααααααααααααααααααααααααααααααααααα→

␈↓"␈↓ α←␈↓∧                              VALIDITY




␈↓"␈↓ α←␈↓α␈↓ αfFig. 7-4.    Expressiveness and Validity of KS Invocation Mechanisms.    
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    227␈↓

␈↓"β␈↓ α←␈↓␈↓αGeneralized invocation criteria␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?All␈αof␈αthe␈αdevelopment␈αthus␈αfar␈α
has␈αbeen␈αin␈αthe␈αcontext␈αof␈α``tuning''␈α
a
␈↓ α←␈↓single␈αcontrol␈αstructure,␈αin␈αthis␈αcase,␈αgoal-directed␈αinvocation.␈α We␈αnoted␈αthat
␈↓ α←␈↓the␈αexpressiveness␈αof␈αcontent-directed␈αinvocation␈αmade␈αpossible␈αtuning␈α
based
␈↓ α←␈↓on␈αa␈αvery␈αwide␈α
range␈αof␈αcriteria: ␈αany␈αcriterion␈α
that␈αcould␈αbe␈αexpressed␈α
in␈αa
␈↓ α←␈↓computable␈α∂predicate␈α∂that␈α⊂examined␈α∂the␈α∂KS␈α⊂code.␈α∂ This␈α∂is␈α⊂one␈α∂interesting
␈↓ α←␈↓ability␈α∪that␈α∪arises␈α∀from␈α∪the␈α∪use␈α∪in␈α∀meta-rules␈α∪of␈α∪an␈α∀explicit,␈α∪functional
␈↓ α←␈↓specification␈αof␈α
desired␈αKSs.␈α That␈α
is,␈αthe␈αmeta-rule␈α
premise␈αis,␈αin␈α
effect,␈αan
␈↓ α←␈↓executable procedure for selecting the appropriate KSs.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
ability␈αto␈α
specify␈αretrieval␈α
criteria␈αcan␈α
be␈αturned␈α
to␈α
another␈αuse,
␈↓ α←␈↓which␈α
we␈α
term␈α
␈↓↓generalized␈α
invocation␈α
criteria␈↓.␈α
 The␈α
point␈α
is␈α
simply␈α∞to␈α
apply
␈↓ α←␈↓this␈α∀criteria-defining␈α∀capability␈α∀to␈α∃the␈α∀specification␈α∀of␈α∀the␈α∃basic␈α∀control
␈↓ α←␈↓regime,␈α∞rather␈α
than␈α∞limiting␈α∞its␈α
use␈α∞to␈α∞an␈α
established␈α∞control␈α∞structure.␈α
 For
␈↓ α←␈↓example, goal-directed invocation might be specified as

␈↓"β␈↓ α←␈↓	␈↓ β∨Use rules that mention the current goal in their action

␈↓ α←␈↓while data-directed invocation might be phrased as

␈↓"β␈↓ α←␈↓	␈↓ α←Use rules that mention the current conclusion in their premise.

␈↓"β␈↓ α←␈↓␈↓ β?We␈α
might␈α
also␈αspecify␈α
new␈α
kinds␈α
of␈αcontrol␈α
regimes,␈α
as,␈α
for␈αinstance,
␈↓ α←␈↓the␈α∀``speed-directed''␈α∪retrieval␈α∀criterion␈α∀mentioned␈α∪above.␈α∀ The␈α∀ability␈α∪to
␈↓ α←␈↓specify␈α⊃such␈α⊂generalized␈α⊃criteria␈α⊃appears␈α⊂to␈α⊃offer␈α⊂at␈α⊃least␈α⊃two␈α⊂advantages.
␈↓ α←␈↓First,␈αit␈αfrees␈αthe␈αprogrammer␈αfrom␈αthe␈αlimitation␈αof␈αusing␈αonly␈αthose␈αcontrol
␈↓ α←␈↓regimes␈α∀(e.g.,␈α∀goal-directed␈α∀or␈α∀data-directed)␈α∀already␈α∀hardwired␈α∃into␈α∀the
␈↓ α←␈↓programming␈αlanguage␈αin␈αuse.␈α Second,␈αit␈αmakes␈αpossible␈αan␈αadded␈αdegree␈α
of
␈↓ α←␈↓explicitness in program representation.
␈↓"β␈↓ α←␈↓␈↓ β?To␈αsee␈α
the␈αbenefits␈α
that␈αaccrue␈αfrom␈α
such␈αexplicitness,␈α
consider␈αwhat
␈↓ α←␈↓typically␈α⊗happens␈α⊗when␈α⊗the␈α↔available␈α⊗set␈α⊗of␈α⊗invocation␈α↔mechanisms␈α⊗is
␈↓ α←␈↓incomplete␈α∪for␈α∀a␈α∪particular␈α∀problem:  ␈α∪The␈α∀programmer␈α∪often␈α∀resorts␈α∪to
␈↓ α←␈↓various␈α
devious␈α
implicit␈αor␈α
indirect␈α
techniques␈αto␈α
achieve␈α
the␈α
desired␈αeffect.
␈↓ α←␈↓One␈αpopular␈αapproach␈αis␈αthat␈αof␈αgetting␈αa␈αmultitude␈αof␈αeffects␈αfrom␈αa␈αsingle
␈↓ α←␈↓mechanism.␈α∪ Where␈α∪KSs␈α∩are␈α∪retrieved␈α∪via␈α∩pre-computed␈α∪index␈α∪lists,␈α∩for
␈↓ α←␈↓instance,␈α∞a␈α∂common␈α∞approach␈α∞is␈α∂to␈α∞hand-tool␈α∞the␈α∂ordering␈α∞of␈α∞these␈α∂lists␈α∞to
␈↓ α←␈↓achieve␈α⊗effects␈α⊗not␈α⊗otherwise␈α∃available␈α⊗in␈α⊗the␈α⊗existing␈α⊗formalism.␈α∃ For
␈↓ α←␈↓example,␈α∃where␈α∃goal-directed␈α∃invocation␈α∃is␈α∃accomplished␈α∃by␈α∃using␈α∀pre-
␈↓ α←␈↓computed␈αlists␈αof␈αoperators,␈αhand-tooling␈αthese␈αlists␈αcan␈αadd␈αa␈αrange␈αof␈α
other
␈↓ α←␈↓control␈α⊂regimes.␈α⊃ In␈α⊂[Waldinger74],␈α⊃for␈α⊂instance,␈α⊃the␈α⊂␈↓	GOALCLASS␈↓␈α⊃lists␈α⊂were
␈↓ α←␈↓hand-ordered␈α∩to␈α⊃insure␈α∩that␈α⊃the␈α∩fastest␈α⊃operators␈α∩were␈α⊃invoked␈α∩first;␈α⊃the
␈↓ α←␈↓analogous␈α∪lists␈α∪in␈α∪␈↓¬MYCIN␈↓␈α∪have,␈α∪in␈α∪the␈α∪past,␈α∪been␈α∪hand-tooled␈α∪to␈α∪effect␈α∩a
␈↓ α←␈↓number of different partial orderings on the rules that are invoked.
␈↓"β␈↓ α←␈↓␈↓ β?A␈αsimilar␈αexample␈αarises␈αwhen␈αusing␈αa␈αmultiple␈αpriority␈αlevel␈αagenda
␈↓ α←␈↓of␈α∂the␈α∂sort␈α∂described␈α∂in␈α∂[Bobrow77].␈α∂ Suppose,␈α∂for␈α∂example,␈α∂we␈α⊂wanted␈α∂to
␈↓ α←␈↓insure␈α∂a␈α⊂particular␈α∂partial␈α⊂ordering␈α∂of␈α∂processes␈α⊂to␈α∂be␈α⊂put␈α∂on␈α⊂the␈α∂agenda.
␈↓ α←␈↓␈↓228    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓Note␈αthat␈αthere␈αis␈αno␈αway␈αto␈αsay␈αexplicitly,␈α␈↓↓make␈αsure␈αthat␈αthese␈αprocesses␈α(in
␈↓ α←␈↓↓set␈α∞A)␈α∞are␈α∞executed␈α
before␈α∞those␈α∞(in␈α∞set␈α∞B)␈↓.␈α
 Instead,␈α∞we␈α∞have␈α∞to␈α∞be␈α
indirect,
␈↓ α←␈↓and␈α∂could,␈α∞for␈α∂instance,␈α∂assign␈α∞a␈α∂priority␈α∂of␈α∞6␈α∂to␈α∞the␈α∂rules␈α∂in␈α∞set␈α∂A␈α∂and␈α∞a
␈↓ α←␈↓priority of 5 to those in B.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∞are␈α∞a␈α∞number␈α∞of␈α∞problems␈α∞associated␈α∞with␈α∞trying␈α∞to␈α∂do␈α∞these
␈↓ α←␈↓sorts␈α⊂of␈α⊃indirect␈α⊂encodings,␈α⊃all␈α⊂of␈α⊃which␈α⊂seem␈α⊃to␈α⊂arise␈α⊃from␈α⊂the␈α⊃fact␈α⊂that
␈↓ α←␈↓information␈αis␈αunavoidably␈αlost␈αby␈αthe␈αindirection␈αinvolved.␈α Note␈αthat␈αin␈αall
␈↓ α←␈↓the␈αcases␈α
above,␈αthe␈α
intent␈αof␈αthe␈α
hand-tooling␈αand␈α
indirect␈αpriority␈αsetting␈α
is
␈↓ α←␈↓nowhere recorded.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
resulting␈α
system␈α
is␈α
both␈α
opaque␈αand␈α
prone␈α
to␈α
bugs.␈α
 To␈α
see␈αthe
␈↓ α←␈↓opacity␈αof␈αthe␈αsystem,␈αconsider,␈αfor␈αinstance,␈αthe␈αagenda␈αexample␈αwhere,␈αafter
␈↓ α←␈↓the␈α
priorities␈α
have␈α∞been␈α
set,␈α
it␈α∞will␈α
not␈α
be␈α
apparent␈α∞␈↓↓why␈↓␈α
the␈α
processes␈α∞in␈α
A
␈↓ α←␈↓were␈α∂given␈α∂higher␈α∂priority␈α⊂than␈α∂those␈α∂in␈α∂B.␈α⊂ Were␈α∂they␈α∂more␈α∂likely␈α⊂to␈α∂be
␈↓ α←␈↓useful␈αor␈αwas␈αit␈αdesirable␈αthat␈αthose␈αin␈αA␈αprecede␈αthose␈αin␈αB␈αno␈αmatter␈αhow
␈↓ α←␈↓useful␈α∂each␈α∂might␈α∂be? ␈α∂After␈α∂a␈α∞while,␈α∂even␈α∂the␈α∂programmer␈α∂who␈α∂set␈α∞these
␈↓ α←␈↓priorities may forget what motivated the particular priorities chosen.
␈↓"β␈↓ α←␈↓␈↓ β?Bugs␈αcan␈α
arise␈αin␈α
this␈αsetting␈α
due␈αboth␈α
to␈αexecution-time␈α
events␈αand
␈↓ α←␈↓events␈αin␈αthe␈α
long-term␈αdevelopment␈αof␈α
the␈αprogram.␈α Consider␈α
for␈αinstance
␈↓ α←␈↓what␈αhappens␈α
if,␈αduring␈αa␈α
run,␈αbefore␈αany␈α
of␈αthe␈αprocesses␈α
in␈αA␈αare␈α
invoked,
␈↓ α←␈↓an␈α
event␈α
occurs␈α
which␈α
makes␈α
it␈α
clear␈α
that␈α
the␈α
priority␈α
of␈α
processes␈α
in␈α
A␈α
ought
␈↓ α←␈↓to␈α
be␈α
reduced␈α∞(for␈α
reasons␈α
unrelated␈α
to␈α∞the␈α
desired␈α
partial␈α
ordering).␈α∞ If␈α
we
␈↓ α←␈↓adjust␈αonly␈α
the␈αpriority␈αof␈α
those␈αin␈α
A,␈αan␈αexecution-time␈α
bug␈αarises,␈αsince␈α
the
␈↓ α←␈↓desired␈αrelative␈αordering␈αmay␈αbe␈αlost.␈α Yet␈αthere␈αis␈αno␈αrecord␈αof␈αthe␈α
necessary
␈↓ α←␈↓interconnection␈α∞of␈α∂priorities␈α∞to␈α∂remind␈α∞us␈α∞to␈α∂adjust␈α∞all␈α∂of␈α∞them.␈α∂ A␈α∞similar
␈↓ α←␈↓problem␈α∞can␈α∞arise␈α
during␈α∞the␈α∞long-term␈α∞development␈α
of␈α∞the␈α∞program␈α∞if␈α
we
␈↓ α←␈↓attempt␈αto␈α
introduce␈αanother␈αindirect␈α
effect␈αby␈αjuggling␈α
priorities␈αand␈αend␈α
up
␈↓ α←␈↓modifying␈αthose␈αin␈αset␈αA␈αwithout␈αmaking␈αthe␈αnecessary␈αadjustments␈αto␈αthose
␈↓ α←␈↓in B.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∪problem␈α∪with␈α∀this␈α∪approach␈α∪is␈α∪that␈α∀it␈α∪tries␈α∪to␈α∪use␈α∀a␈α∪single
␈↓ α←␈↓invocation␈α⊂mechanism␈α⊂to␈α⊂accomplish␈α⊂a␈α∂number␈α⊂of␈α⊂effects.␈α⊂ It␈α⊂does␈α⊂this␈α∂by
␈↓ α←␈↓reducing␈α
a␈αnumber␈α
of␈α
different,␈αincommensurate␈α
factors␈αto␈α
a␈α
single␈αnumber,
␈↓ α←␈↓␈↓↓with␈α⊂no␈α⊂record␈α⊂of␈α⊂how␈α⊂that␈α⊂number␈α⊂was␈α⊂reached␈↓.␈α⊂ Meta-rules,␈α⊂on␈α⊂the␈α∂other
␈↓ α←␈↓hand,␈α⊂offer␈α⊂a␈α∂mechanism␈α⊂for␈α⊂making␈α∂these␈α⊂sorts␈α⊂of␈α⊂considerations␈α∂explicit
␈↓ α←␈↓and␈α
for␈αleaving␈α
a␈α
record␈αof␈α
why␈α
a␈αset␈α
of␈α
processes␈αwas␈α
queued␈α
in␈αa␈α
particular
␈↓ α←␈↓order.
␈↓"β␈↓ α←␈↓␈↓ β?Meta-rules␈α⊂also␈α∂offer␈α⊂the␈α⊂advantage␈α∂of␈α⊂␈↓↓localizing␈↓␈α∂all␈α⊂of␈α⊂the␈α∂control
␈↓ α←␈↓information.␈α
 Note␈α∞that␈α
juggling␈α∞priorities␈α
means␈α
trying␈α∞to␈α
achieve␈α∞a␈α
global
␈↓ α←␈↓effect␈α∪via␈α∪a␈α∪number␈α∪of␈α∩scattered␈α∪local␈α∪adjustments.␈α∪ This␈α∪is␈α∪often␈α∩quite
␈↓ α←␈↓difficult␈α∞and␈α∞can␈α∞be␈α∞very␈α∂hard␈α∞to␈α∞change␈α∞or␈α∞update.␈α∞ Localizing␈α∂each␈α∞such
␈↓ α←␈↓invocation␈α⊃criterion␈α⊂in␈α⊃a␈α⊂single␈α⊃meta-rule␈α⊂makes␈α⊃subsequent␈α⊂modifications
␈↓ α←␈↓easier.␈α Since␈αall␈αof␈αthe␈αinformation␈αis␈αin␈αone␈αplace,␈αchanging␈αa␈αcriterion␈αcan
␈↓ α←␈↓be␈α∩accomplished␈α⊃by␈α∩editing␈α⊃the␈α∩relevant␈α⊃meta-rule␈α∩rather␈α∩than␈α⊃searching
␈↓ α←␈↓through␈α∂a␈α∂program␈α∂for␈α∂all␈α∂the␈α∂places␈α∂in␈α∂which␈α∂priorities␈α∂have␈α∂been␈α∂set␈α∞to
␈↓ α←␈↓effect that criterion.
␈↓"β␈↓ α←␈↓␈↓ β?Our␈α⊗historical␈α⊗overview␈α∃provides␈α⊗another␈α⊗observation.␈α∃ Consider
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    229␈↓

␈↓"β␈↓ α←␈↓viewing␈α∪the␈α∪progress␈α∀from␈α∪standard␈α∪procedure-calling␈α∪to␈α∀techniques␈α∪like
␈↓ α←␈↓goal-directed␈α⊗invocation␈α↔as␈α⊗making␈α↔it␈α⊗possible␈α⊗to␈α↔be␈α⊗less␈α↔precise,␈α⊗and
␈↓ α←␈↓therefore␈α
less␈α
restrictive,␈α∞about␈α
the␈α
role␈α
of␈α∞a␈α
given␈α
procedure␈α
in␈α∞a␈α
program.
␈↓ α←␈↓Where␈α
invocation␈α
by␈α
name␈α
(i.e.,␈α
procedures)␈α
requires␈α
that␈α
we␈α∞decide␈α
exactly
␈↓ α←␈↓␈↓↓where␈↓␈αand␈α␈↓↓when␈↓␈αthe␈αcode␈αis␈αto␈αbe␈αinvoked,␈αgoal-directed␈α
invocation␈αrequires
␈↓ α←␈↓only␈α∃that␈α∀we␈α∃specify␈α∀␈↓↓how␈↓␈α∃a␈α∀procedure␈α∃is␈α∀to␈α∃be␈α∀used.␈α∃ The␈α∀perspective
␈↓ α←␈↓suggested␈α
here␈αmoves␈α
us␈α
another␈αstep␈α
along␈α
that␈αline,␈α
by␈αconsidering␈α
retrieval
␈↓ α←␈↓separately␈α∂from␈α∂the␈α∂writing␈α∂of␈α⊂a␈α∂KS.␈α∂ The␈α∂programmer␈α∂can,␈α∂if␈α⊂he␈α∂desires,
␈↓ α←␈↓write␈αa␈αKS␈αwithout␈αspecifying␈αhow␈αit␈αis␈αto␈αbe␈αused␈αand␈αcan␈αleave␈αthis␈α
up␈αto
␈↓ α←␈↓the invocation criteria to decide.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∞point␈α∞here␈α∞is␈α∞not␈α
that␈α∞the␈α∞programmer␈α∞should␈α∞not␈α∞specify␈α
such
␈↓ α←␈↓information,␈αsince␈αit␈αcan␈αbe␈αvery␈αuseful␈αfor␈αcutting␈αdown␈αcombinatorics.␈α For
␈↓ α←␈↓inference␈α
rules,␈α
for␈αinstance,␈α
it␈α
is␈αimportant␈α
to␈α
know␈αwhich␈α
are␈α
more␈αuseful␈α
in
␈↓ α←␈↓the␈α∩goal-directed␈α∪mode␈α∩and␈α∪which␈α∩are␈α∩more␈α∪useful␈α∩in␈α∪the␈α∩data-directed
␈↓ α←␈↓mode.␈α⊃ Such␈α∩information␈α⊃can␈α∩cut␈α⊃down␈α∩search␈α⊃or␈α∩control␈α⊃the␈α∩number␈α⊃of
␈↓ α←␈↓forward␈α⊂inferences␈α∂drawn␈α⊂from␈α∂a␈α⊂new␈α⊂assertion.␈α∂ The␈α⊂point␈α∂is␈α⊂that␈α⊂if␈α∂the
␈↓ α←␈↓programmer␈αdoes␈αwant␈αto␈αspecify␈αsuch␈αthings,␈αit␈αis␈αbetter␈αif␈αhe␈αis␈αnot␈αlimited
␈↓ α←␈↓to␈αmaking␈αthat␈αinformation␈αsynonymous␈αwith␈αthe␈αhandle␈αused␈αto␈αretrieve␈αthe
␈↓ α←␈↓code.␈α⊂ (This␈α⊃occurred␈α⊂in␈α⊂␈↓¬PLANNER␈↓,␈α⊃for␈α⊂instance,␈α⊂where␈α⊃the␈α⊂indication␈α⊃that␈α⊂a
␈↓ α←␈↓theorem␈α∞ought␈α∞to␈α∞be␈α∞used␈α∞in␈α∞a␈α∞goal-directed␈α∞mode␈α∞for␈α∞a␈α∞particular␈α∞pattern
␈↓ α←␈↓became␈αthe␈αsole␈αway␈αof␈αindexing␈α
it␈α[by␈αthe␈αsingle␈αgoal␈αpattern].) ␈αIf␈α
we␈αkeep
␈↓ α←␈↓these␈α⊃two␈α⊂things␈α⊃separate,␈α⊃we␈α⊂allow␈α⊃information␈α⊂about␈α⊃appropriate␈α⊃use␈α⊂to
␈↓ α←␈↓become a piece of advice rather than a constraint.

␈↓"β␈↓ α←␈↓␈↓αFlexibility␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂choice␈α∂of␈α∂a␈α∂KS␈α∂reference␈α∂technique␈α∂(name,␈α∂external␈α∞descriptor,
␈↓ α←␈↓content␈α
reference)␈αcan␈α
have␈αa␈α
significant␈αimpact␈α
on␈αthe␈α
difficulty␈α
of␈αmaking
␈↓ α←␈↓changes␈αto␈α
a␈αprogram.␈α
 As␈αwe␈α
will␈αsee,␈α
content␈αreference␈α
offers␈αa␈α
number␈αof
␈↓ α←␈↓advantages in this respect.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α%consideration␈α%becomes␈α%particularly␈α%important␈α%for␈α%the
␈↓ α←␈↓applications␈α∩we␈α∩have␈α∪described: ␈α∩Since␈α∩much␈α∪of␈α∩our␈α∩effort␈α∪has␈α∩involved
␈↓ α←␈↓making␈αit␈αpossible␈αfor␈αthe␈αexpert␈αto␈αaugment␈αthe␈αknowledge␈αbase,␈αwe␈αshould
␈↓ α←␈↓take␈α∀advantage␈α∀of␈α∀any␈α∃means␈α∀of␈α∀minimizing␈α∀the␈α∀difficulty␈α∃involved␈α∀in
␈↓ α←␈↓propagating␈α∃the␈α∃effects␈α∃of␈α∃a␈α∀change.␈α∃ Such␈α∃flexibility␈α∃will␈α∃also␈α∀become
␈↓ α←␈↓increasingly␈α∂important␈α⊂as␈α∂knowledge␈α∂base␈α⊂construction␈α∂proceeds,␈α∂since␈α⊂as␈α∂a
␈↓ α←␈↓program␈α
gets␈αlarger␈α
it␈α
becomes␈αincreasingly␈α
difficult␈α
to␈αcope␈α
with␈α
the␈αeffects
␈↓ α←␈↓of changes.
␈↓"β␈↓ α←␈↓␈↓ β?For␈αthe␈αsake␈αof␈αdiscussion,␈αwe␈αidentify␈αtwo␈αclasses␈αof␈αsuch␈αflexibility: 
␈↓ α←␈↓␈↓↓Compile␈αtime␈αflexibility␈↓␈αis␈αthe␈αability␈αto␈αmake␈αchanges␈αin␈αthe␈αknowledge␈αbase
␈↓ α←␈↓between␈α≤performance␈α≤runs␈α≤and␈α≠then␈α≤have␈α≤those␈α≤changes␈α≠integrated
␈↓ α←␈↓throughout␈αthe␈αsystem;␈α␈↓↓execution␈αtime␈αflexibility␈↓␈αrefers␈αto␈αthe␈αability␈αto␈αswitch
␈↓ α←␈↓strategies␈α
during␈α
execution,␈α
under␈α
program␈α
control.␈α
 This␈α
section␈α∞deals␈α
with
␈↓ α←␈↓the former.  Comments in Section 7-5-5 below deal with the latter.
␈↓"β␈↓ α←␈↓␈↓ β?To␈α⊃judge␈α⊃the␈α⊃impact␈α⊃of␈α∩selecting␈α⊃one␈α⊃or␈α⊃another␈α⊃of␈α∩the␈α⊃reference
␈↓ α←␈↓techniques,␈α∞we␈α∞examine␈α
two␈α∞types␈α∞of␈α∞changes: ␈α
editing␈α∞or␈α∞adding␈α∞a␈α
(object-
␈↓ α←␈↓␈↓230    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓level)␈α
KS␈α∞to␈α
the␈α
system,␈α∞and␈α
editing␈α
or␈α∞adding␈α
a␈α
strategy␈α∞(meta-rule).␈↓
19␈↓␈α
See
␈↓ α←␈↓Table 7-2 for an overview.
␈↓"β␈↓ α←␈↓␈↓ β?Consider,␈α⊂for␈α⊂example,␈α⊂the␈α⊂effect␈α∂of␈α⊂editing␈α⊂(or␈α⊂adding)␈α⊂a␈α⊂KS␈α∂and
␈↓ α←␈↓imagine␈α⊃that␈α⊃strategies␈α⊃used␈α⊂the␈α⊃reference-by-name␈α⊃approach.␈α⊃ First,␈α⊂after
␈↓ α←␈↓editing␈αa␈α
KS␈αin␈α
such␈αa␈αsystem,␈α
all␈αstrategies␈α
that␈αmention␈α
it␈αmust␈αbe␈α
retrieved
␈↓ α←␈↓and␈αexamined␈αto␈αsee␈αif␈αthey␈αstill␈αapply,␈αand␈αthen␈αmust␈αbe␈αedited␈αaccordingly.
␈↓ α←␈↓Next,␈αsince␈α
it␈αis␈αalso␈α
possible␈αthat␈α
the␈αrevised␈αKS␈α
should␈αnow␈α
be␈αmentioned
␈↓ α←␈↓in␈α∞other␈α∞strategies,␈α∞the␈α∞rest␈α∞of␈α∞the␈α∞strategies␈α∞must␈α∞also␈α∞be␈α∞examined.␈α
 Using
␈↓ α←␈↓the␈α∃external-descriptor␈α∃approach,␈α∃we␈α∃need␈α∃only␈α∃update␈α⊗the␈α∃appropriate
␈↓ α←␈↓descriptors,␈α
which␈α∞would␈α
be␈α∞stored␈α
with␈α∞the␈α
KS.␈α∞ In␈α
addition,␈α∞the␈α
updating
␈↓ α←␈↓required␈α∞should␈α∞be␈α∞evident␈α∞from␈α∞the␈α
sort␈α∞of␈α∞editing␈α∞done␈α∞on␈α∞the␈α∞KS␈α
itself.
␈↓ α←␈↓All␈αrelevant␈αstrategies␈αwill␈α
then␈αautomatically␈αadjust␈αto␈αthese␈α
changes.␈α With
␈↓ α←␈↓content␈α∞reference␈α∞there␈α∞is␈α∞no␈α∞additional␈α∞effort␈α∞of␈α∞even␈α∂updating␈α∞descriptors
␈↓ α←␈↓since the strategies will adjust to the changes found in the edited KS.
␈↓"β␈↓ α←␈↓␈↓ β?Adding␈α⊂a␈α⊂new␈α⊂strategy␈α∂to␈α⊂the␈α⊂system␈α⊂(or␈α∂revising␈α⊂an␈α⊂old␈α⊂one)␈α∂also
␈↓ α←␈↓causes␈α∀problems␈α∀for␈α∀the␈α∀reference-by-name␈α∀approach:␈α∀It␈α∀is␈α∃necessary␈α∀to
␈↓ α←␈↓review␈α∂all␈α∂the␈α∞KSs␈α∂to␈α∂determine␈α∞which␈α∂the␈α∂new␈α∞or␈α∂revised␈α∂strategy␈α∞should
␈↓ α←␈↓mention.␈↓
20␈↓␈α
 ␈αUsing␈α
external␈α
descriptors,␈αit␈α
is␈α
possible␈αthat␈α
no␈αadditional␈α
effort
␈↓ α←␈↓is␈α⊗required,␈α↔if␈α⊗the␈α↔description␈α⊗in␈α↔the␈α⊗new␈α↔strategy␈α⊗uses␈α↔the␈α⊗available
␈↓ α←␈↓``vocabulary''␈α∞of␈α∞descriptors.␈α∞ If,␈α∞however,␈α∞it␈α∞requires␈α∞a␈α∞descriptor␈α∞not␈α∞yet␈α∞in
␈↓ α←␈↓that␈α
vocabulary,␈α
we␈α
have␈α
the␈α
formidable␈α
task␈α
of␈α
reviewing␈α
all␈α∞existing␈α
KSs
␈↓ α←␈↓and adding to each the appropriate entry for the new descriptor.
␈↓"β␈↓ α←␈↓␈↓ β?Thus␈α∞there␈α∞is␈α∞a␈α∞fundamental␈α∞shortcoming␈α∞in␈α∞the␈α∞external␈α∞descriptor
␈↓ α←␈↓approach␈αbecause␈αit␈αuses␈αa␈αfixed␈αnumber␈αof␈αdescriptive␈αterms.␈α Since␈αadding
␈↓ α←␈↓a␈α
new␈α
term␈α
to␈α
this␈α
set␈α
can␈α
involve␈α
a␈α
lot␈α
of␈α
work,␈α
it␈α
becomes␈α
a␈α
task␈αthat␈α
should
␈↓ α←␈↓not␈α∀be␈α∀undertaken␈α∪very␈α∀often.␈α∀ Avoiding␈α∪this␈α∀updating␈α∀problem␈α∀is␈α∪one
␈↓ α←␈↓important␈α
advantage␈α
of␈α
content␈α∞reference: ␈α
It␈α
gives␈α
meta-rules␈α
the␈α∞ability␈α
to
␈↓ α←␈↓``go␈α∞in␈α
and␈α∞look''␈α
for␈α∞any␈α
characteristic␈α∞deemed␈α
significant.␈α∞ As␈α
a␈α∞result,␈α
the
␈↓ α←␈↓addition␈α∞of␈α∞a␈α
new␈α∞strategy␈α∞with␈α
a␈α∞new␈α∞meta-level␈α
conceptual␈α∞primitive␈α∞is␈α
a
␈↓ α←␈↓transparent operation.
␈↓"β␈↓ α←␈↓␈↓ β?To␈α∪make␈α∪this␈α∪clear,␈α∩consider␈α∪the␈α∪following␈α∪simple␈α∪example.␈α∩ The
␈↓ α←␈↓performance␈α⊂program's␈α⊂backward␈α⊂chaining␈α⊂of␈α⊂rules␈α⊂produces␈α⊂a␈α∂depth-first

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[19]␈αWhile␈αthe␈αimpact␈αis␈αillustrated␈αin␈αthe␈αcontext␈αof␈αmeta-rules␈α
and␈αobject-
␈↓ α←␈↓level␈α∂rules,␈α∂the␈α∂point␈α∞is␈α∂more␈α∂generally␈α∂applicable␈α∂to␈α∞any␈α∂kind␈α∂of␈α∂KS␈α∂in␈α∞a
␈↓ α←␈↓system invoking any other KS, regardless of the level of each.

␈↓"β␈↓ α←␈↓[20]␈α∂There␈α∂is␈α∞a␈α∂plausible␈α∂objection␈α∂to␈α∞this: ␈α∂It␈α∂may␈α∞be␈α∂claimed␈α∂that␈α∂a␈α∞new
␈↓ α←␈↓strategy␈α
is␈αoften␈α
written␈α
with␈αsome␈α
specific␈α
circumstance␈αand␈α
purpose␈αin␈α
mind
␈↓ α←␈↓and␈α
that␈αthis␈α
clearly␈αrestricts␈α
the␈αnumber␈α
of␈α
KSs␈αthat␈α
need␈αbe␈α
considered␈αto␈α
a
␈↓ α←␈↓small␈α∂subset␈α⊂of␈α∂the␈α∂total.␈α⊂ This␈α∂is␈α∂entirely␈α⊂correct.␈α∂ And␈α∂it␈α⊂is␈α∂on␈α⊂just␈α∂such
␈↓ α←␈↓grounds␈αthat␈αwe␈αwould␈αstart␈αto␈αbuild␈αthe␈αset␈αof␈αdescriptors␈αto␈αbe␈αused␈αin␈αthe
␈↓ α←␈↓reference␈α∞by␈α
description␈α∞approach.␈α∞ In␈α
part,␈α∞then,␈α∞the␈α
technique␈α∞is␈α∞simply␈α
a
␈↓ α←␈↓step␈αtoward␈αgreater␈αformalization␈αof␈αknowledge␈αalready␈αused␈αin␈αinformal␈α
and
␈↓ α←␈↓ad hoc ways.
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    231␈↓

␈↓"β␈↓ α←␈↓search␈α⊃of␈α⊃an␈α⊃and/or␈α⊃goal␈α⊃tree,␈α⊃where␈α⊃each␈α⊃clause␈α⊃of␈α⊃a␈α⊃rule␈α⊃may␈α⊂possibly
␈↓ α←␈↓sprout␈αa␈α
new␈αsub-tree.␈α It␈α
might␈αbe␈αuseful,␈α
then,␈αto␈αtry␈α
rules␈αwith␈α
the␈αfewest
␈↓ α←␈↓clauses␈α∂first.␈α∂ This␈α⊂approach␈α∂would␈α∂require␈α⊂a␈α∂strategy␈α∂that␈α⊂said␈α∂something
␈↓ α←␈↓like␈α
␈↓↓try␈α
those␈α
rules␈α
with␈α
three␈α
or␈α
fewer␈α
premise␈α
clauses␈α
first␈↓,␈α
and␈αmight␈α
require
␈↓ α←␈↓a␈α∪new␈α∪meta-level␈α∀primitive.␈α∪ In␈α∪the␈α∀external␈α∪descriptor␈α∪case,␈α∀a␈α∪property
␈↓ α←␈↓indicating␈αthe␈αrelevant␈αinformation␈αwould␈αhave␈αto␈αbe␈αadded␈αto␈α
every␈αrule.␈↓
21␈↓
␈↓ α←␈↓Content␈αreference␈αmakes␈αthe␈αtask␈αmuch␈αeasier: ␈αWe␈αcan␈αwrite␈αa␈αnew␈αfunction
␈↓ α←␈↓that␈αcounts␈αthe␈αnumber␈αof␈αpremise␈αclauses␈αfound␈αin␈αan␈αobject-level␈αrule␈αand
␈↓ α←␈↓use␈αthe␈αfunction␈αin␈αa␈αmeta-rule.␈α Nothing␈αat␈αall␈αneed␈αbe␈αdone␈αto␈αany␈αobject-
␈↓ α←␈↓level rule.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α
claim,␈α
in␈α
addition,␈α
that␈α
the␈α
set␈α
of␈α
useful␈α
meta-level␈α
primitives␈α
is
␈↓ α←␈↓potentially␈α∂large␈α∂and␈α∞therefore␈α∂difficult␈α∂to␈α∂define␈α∞a␈α∂priori.␈α∂ Thus,␈α∂over␈α∞the
␈↓ α←␈↓course␈αof␈αmost␈αprogram␈αdevelopment␈αperiods,␈αit␈αis␈αeffectively␈αan␈αopen␈α
set,␈αto
␈↓ α←␈↓which␈α⊂new␈α∂members␈α⊂are␈α⊂continually␈α∂being␈α⊂added.␈α⊂ It␈α∂is␈α⊂thus␈α⊂important␈α∂to
␈↓ α←␈↓make␈α
this␈α
task␈α
as␈α
easy␈αas␈α
possible.␈α
 Where␈α
the␈α
external␈α
descriptor␈αapproach
␈↓ α←␈↓requires␈α∂analyzing␈α∂each␈α∂new␈α∂KS␈α∂with␈α∂reference␈α∂to␈α∂each␈α∂of␈α∂the␈α∂descriptors,
␈↓ α←␈↓content␈αreference␈αrequires␈αsimply␈αthat␈αthe␈αnew␈αKS␈αbe␈α``added␈αto␈αthe␈αpot.''  ␈αIt
␈↓ α←␈↓will␈α∨subsequently␈α∨be␈α≡referenced␈α∨by␈α∨any␈α≡strategy␈α∨that␈α∨describes␈α≡it
␈↓ α←␈↓appropriately.
␈↓"β␈↓ α←␈↓␈↓ β?To␈α⊂summarize,␈α⊃consider␈α⊂the␈α⊂difference␈α⊃between␈α⊂the␈α⊂first␈α⊃and␈α⊂third
␈↓ α←␈↓rows␈α
of␈α
Table␈α
7-2.␈α
 Note␈α
in␈α
particular␈α
how␈α
much␈α
easier␈α
it␈α
is␈α
to␈α
accomplish
␈↓ α←␈↓a␈α∪number␈α∪of␈α∪standard␈α∀knowledge␈α∪base␈α∪modifications.␈α∪ This␈α∪can␈α∀offer␈α∪a
␈↓ α←␈↓substantive advantage when the knowledge base becomes appreciably large.















␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[21]␈α
This␈α
may␈α
not␈α∞be␈α
a␈α
very␈α
difficult␈α∞task␈α
if␈α
it␈α
is␈α∞possible␈α
to␈α
write␈α
a␈α∞bit␈α
of
␈↓ α←␈↓code␈αthat␈αcomputes␈αthe␈αrelevant␈αinformation␈αand␈αthen␈αapply␈αit␈αto␈αevery␈αrule.
␈↓ α←␈↓This␈α∪is,␈α∪in␈α∪fact,␈α∪what␈α∪we␈α∪are␈α∪suggesting,␈α∪with␈α∪two␈α∪small␈α∀but␈α∪important
␈↓ α←␈↓differences.␈α⊃ First,␈α⊂we␈α⊃don't␈α⊂bother␈α⊃to␈α⊂compute␈α⊃all␈α⊂properties␈α⊃of␈α⊃all␈α⊂object
␈↓ α←␈↓rules,␈αso␈αa␈αconsiderable␈αamount␈αof␈αspace␈αmay␈αbe␈αsaved.␈α Second,␈αif␈αthis␈αbit␈α
of
␈↓ α←␈↓code␈α
is␈α
kept␈α
around␈α
in␈α
a␈α∞meta-rule,␈α
the␈α
addition␈α
of␈α
a␈α
new␈α∞object-level␈α
rule
␈↓ α←␈↓later on is much easier.
␈↓ α←␈↓␈↓232    STRATEGIES␈↓ 
#7-5␈↓







␈↓ α←␈↓α␈↓ ∧LTable 7-2.    Flexibility Benchmarks.    


␈↓"␈↓ α←␈↓∧ααααααααααααααααπααααααααααααααααααααααπααααααααααααααααααααααα
␈↓"␈↓ α←␈↓∧ REFERENCE      ~    Edit or add       ~  Edit or add
␈↓"␈↓ α←␈↓∧ TECHNIQUE      ~    object-level KS   ~  strategy
␈↓"␈↓ α←␈↓∧ααααααααααααααααβααααααααααααααααααααααβααααααααααααααααααααααα
␈↓"␈↓ α←␈↓∧                ~                      ~
␈↓"␈↓ α←␈↓∧                ~    Check all         ~  Check all KSs
␈↓"␈↓ α←␈↓∧ Reference by   ~    strategies to     ~  to see which it
␈↓"␈↓ α←␈↓∧ name           ~    see which should  ~  should name
␈↓"␈↓ α←␈↓∧                ~    name it           ~
␈↓"␈↓ α←␈↓∧                ~                      ~
␈↓"␈↓ α←␈↓∧                ~                      ~
␈↓"␈↓ α←␈↓∧ Reference by   ~                      ~  Possibly no
␈↓"␈↓ α←␈↓∧ description    ~    Update its        ~  additional effort,
␈↓"␈↓ α←␈↓∧ via external   ~    descriptors       ~  possibly have to add
␈↓"␈↓ α←␈↓∧ descriptors    ~                      ~  new descriptor
␈↓"␈↓ α←␈↓∧                ~                      ~
␈↓"␈↓ α←␈↓∧                ~                      ~
␈↓"␈↓ α←␈↓∧ Reference by   ~                      ~
␈↓"␈↓ α←␈↓∧ description    ~    No additional     ~  No additional effort
␈↓"␈↓ α←␈↓∧ via content    ~    effort            ~
␈↓"␈↓ α←␈↓∧ reference      ~                      ~
␈↓"␈↓ α←␈↓∧αααααααααααααααα∀αααααααααααααααααααααα∀ααααααααααααααααααααααα
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    233␈↓

␈↓"β␈↓ α←␈↓␈↓α7-5-4    Limitations␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂previous␈α∂sections␈α∂noted␈α∂a␈α∞number␈α∂of␈α∂the␈α∂implications␈α∂of␈α∞using
␈↓ α←␈↓content-directed␈α⊃invocation:  ␈α∩increased␈α⊃validity␈α∩and␈α⊃expressiveness␈α∩of␈α⊃the
␈↓ α←␈↓handle␈α∂used␈α∂to␈α∂retrieve␈α∂a␈α∂KS,␈α∂the␈α∂ability␈α∂to␈α∂specify␈α∂generalized␈α∂invocation
␈↓ α←␈↓criteria,␈α∂and␈α∞the␈α∂flexibility␈α∂of␈α∞the␈α∂system␈α∂in␈α∞responding␈α∂to␈α∂changes.␈α∞ There
␈↓ α←␈↓are,␈α∩of␈α∩course,␈α∩a␈α∩number␈α⊃of␈α∩substantive␈α∩difficulties␈α∩in␈α∩realizing␈α∩all␈α⊃these
␈↓ α←␈↓benefits␈α∞and␈α∂a␈α∞number␈α∂of␈α∞limitations␈α∂to␈α∞the␈α∂techniques␈α∞we␈α∂have␈α∞suggested.
␈↓ α←␈↓These␈α⊂are␈α⊂reviewed␈α∂here␈α⊂to␈α⊂help␈α∂outline␈α⊂the␈α⊂range␈α∂of␈α⊂applicability␈α⊂of␈α∂the
␈↓ α←␈↓work.

␈↓"β␈↓ α←␈↓␈↓αValidity␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃validity␈α∩of␈α⊃the␈α∩``handle''␈α⊃provided␈α∩by␈α⊃content␈α∩reference␈α⊃arises
␈↓ α←␈↓from␈αthe␈αability␈αto␈αexamine␈αthe␈αcode␈αof␈αthe␈αKS,␈αyet␈αwe␈αhave␈αdealt␈αsomewhat
␈↓ α←␈↓blithely␈α∞in␈α
previous␈α∞sections␈α∞with␈α
this␈α∞difficult␈α∞problem.␈α
 There␈α∞are␈α∞at␈α
least
␈↓ α←␈↓two␈α⊗sources␈α⊗of␈α↔difficulty.␈α⊗ First,␈α⊗even␈α↔for␈α⊗a␈α⊗programmer,␈α↔reading␈α⊗and
␈↓ α←␈↓understanding␈α
an␈α
arbitrary␈α
chunk␈α
of␈α
code␈α
can␈α
be␈α
difficult.␈α
 Second,␈α
even␈αif
␈↓ α←␈↓we␈α∞could␈α
read␈α∞it,␈α∞it␈α
is␈α∞not␈α∞always␈α
clear␈α∞what␈α
to␈α∞look␈α∞for:␈α
Try␈α∞to␈α∞specify␈α
for
␈↓ α←␈↓instance how to detect by reading its code what goal a procedure achieves.
␈↓"β␈↓ α←␈↓␈↓ β?␈↓¬TEIRESIAS␈↓␈α∪currently␈α∀has␈α∪only␈α∪the␈α∀simplest␈α∪form␈α∀of␈α∪code-examining
␈↓ α←␈↓ability,␈αmade␈α
possible␈αby␈α
several␈αuseful␈α
shortcuts.␈α We␈α
rely␈αfirst␈α
on␈αthe␈α
nature
␈↓ α←␈↓of␈αthe␈α
representation␈αin␈αuse.␈α
 The␈αorganization␈αof␈α
information␈αin␈αrules␈α
makes
␈↓ α←␈↓it␈α∂possible␈α∂to␈α∂say␈α∞that,␈α∂for␈α∂example,␈α∂``␈↓↓rules␈α∞which␈α∂mention␈α∂␈↓	IDENT␈↓↓ity␈α∂in␈α∞their
␈↓ α←␈↓↓conclusion␈↓''␈α∂is␈α∂an␈α⊂example␈α∂of␈α∂goal-directed␈α⊂retrieval.␈α∂ Second,␈α∂the␈α⊂rules␈α∂are
␈↓ α←␈↓viewed␈αas␈αa␈αtask-specific␈αhigh-level␈αlanguage: ␈αThe␈αprimitive␈αterms␈αthey␈αuse
␈↓ α←␈↓are␈α∀both␈α∀domain-specific␈α∀and␈α∀reasonably␈α∀abstract␈α∀(e.g.,␈α∀``␈↓↓the␈α∀culture␈α∀was
␈↓ α←␈↓↓obtained␈α∂from␈α∞a␈α∂␈↓	STERILE␈α∂SOURCE␈↓'').␈α∞ This␈α∂makes␈α∂their␈α∞code␈α∂much␈α∂easier␈α∞to
␈↓ α←␈↓decipher␈αthan,␈αsay,␈αan␈αassembly␈αcode␈αversion␈αof␈αthe␈αsame␈αthing.␈α Finally,␈αthe
␈↓ α←␈↓code␈αis␈αstrongly␈αstylized␈α(the␈αpredicates␈αand␈αassociative␈αtriples),␈αallowing␈αus␈α
to
␈↓ α←␈↓use␈α∀the␈α∀template␈α∪associated␈α∀with␈α∀each␈α∀predicate␈α∪function␈α∀as␈α∀a␈α∀guide␈α∪to
␈↓ α←␈↓deciphering␈α
code.␈α
 As␈α
noted␈α
in␈α∞chapter␈α
2,␈α
the␈α
template␈α
describes␈α∞the␈α
format
␈↓ α←␈↓of␈αa␈αcall␈αto␈αthe␈αfunction␈α(the␈αorder␈αand␈αgeneric␈αtype␈αof␈αits␈αarguments),␈αmuch
␈↓ α←␈↓like␈αa␈αsimplified␈αprocedure␈αdeclaration.␈α Each␈αpredicate␈αfunction␈αthus␈αcarries
␈↓ α←␈↓a␈α∪description␈α∪of␈α∪its␈α∩own␈α∪calls,␈α∪and␈α∪by␈α∩referring␈α∪to␈α∪that␈α∪description␈α∩(i.e.,
␈↓ α←␈↓retrieving␈α
the␈αtemplate␈α
associated␈αwith␈α
the␈α␈↓	CAR␈↓␈α
of␈αa␈α
␈↓¬LISP␈↓␈αform),␈α
we␈αcan␈α
dissect
␈↓ α←␈↓the call into its components (see Section 2-4-4).
␈↓"β␈↓ α←␈↓␈↓ β?We␈αhave,␈αof␈αcourse,␈α
thus␈αfar␈αused␈αthis␈α
ability␈αin␈αonly␈αthe␈α
most␈αbasic,
␈↓ α←␈↓syntactic␈α⊗ways.␈α⊗ We␈α⊗have␈α⊗not␈α∃yet␈α⊗developed␈α⊗means␈α⊗for␈α⊗deriving␈α∃more
␈↓ α←␈↓interesting␈αsemantic␈αcontent␈αby␈αexamining␈αcode,␈αand␈αthis␈αis␈αclearly␈αa␈αdifficult
␈↓ α←␈↓problem.␈α⊃ While␈α⊃the␈α⊂shortcuts␈α⊃described␈α⊃can␈α⊂be␈α⊃used␈α⊃with␈α⊂representations
␈↓ α←␈↓other␈α∞than␈α∞rules,␈α∞they␈α∞are␈α∞not␈α∞universally␈α∞applicable.␈α∞ But␈α∞while␈α∞this␈α
whole
␈↓ α←␈↓problem␈α∞is␈α∞difficult,␈α∞it␈α∞is␈α∞also␈α∂a␈α∞separable␈α∞issue.␈α∞ That␈α∞is,␈α∞the␈α∞extent␈α∂of␈α∞the
␈↓ α←␈↓current␈α⊃capability␈α⊃to␈α⊃examine␈α∩code␈α⊃is␈α⊃extremely␈α⊃elementary,␈α⊃but␈α∩even␈α⊃the
␈↓ α←␈↓simplest␈α∩form␈α∪of␈α∩it␈α∩makes␈α∪available␈α∩the␈α∩interesting␈α∪capabilities␈α∩displayed
␈↓ α←␈↓above.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α⊂more␈α⊃subtle␈α⊂point␈α⊂arises␈α⊃out␈α⊂of␈α⊂the␈α⊃possible␈α⊂suggestion␈α⊃that␈α⊂the
␈↓ α←␈↓␈↓234    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓scheme␈α
proposed␈αhere␈α
appears␈α
to␈αbe␈α
susceptible␈α
to␈αthe␈α
same␈α
criticism␈αmade␈α
of
␈↓ α←␈↓␈↓¬PLANNER␈↓: ␈α
In␈αour␈α
case,␈α
the␈αuser␈α
writes␈α
both␈αthe␈α
KS␈α
code␈αand␈α
the␈αpredicates␈α
that
␈↓ α←␈↓examine␈α
it.␈α
 There␈α
would␈α
seem␈αto␈α
be␈α
room␈α
for␈α
chicanery␈α
in␈αwriting␈α
predicates
␈↓ α←␈↓specially␈αtailored␈αto␈αthe␈αbody␈αof␈αa␈αparticular␈αKS.␈α It␈αis␈αpossible,␈α
for␈αinstance,
␈↓ α←␈↓to␈α
write␈αa␈α
predicate␈αthat␈α
tests␈αfor␈α
the␈αappearance␈α
of␈αa␈α
uniquely␈α
named␈αlocal
␈↓ α←␈↓variable␈αin␈αa␈αprogram␈α
body,␈αor␈αthat␈αperhaps␈αchecks␈α
the␈αtime␈αof␈αday␈αor␈α
phase
␈↓ α←␈↓of the moon.
␈↓"β␈↓ α←␈↓␈↓ β?Our␈α
claim,␈α
however,␈α
was␈α
not␈α
that␈α
the␈α
invocation␈α
criteria␈αwould␈α
always
␈↓ α←␈↓be␈αmeaningful,␈α
but␈αonly␈α
that␈αthey␈α
would␈αbe␈α
made␈αexplicit␈α
and␈αprecise.␈α
 Thus,
␈↓ α←␈↓while␈αit␈αis␈αcurious␈αto␈αask␈αfor␈αa␈αKS␈αthat␈αcontains␈αthe␈αvariable␈α␈↓	ZZ$XX␈↓,␈αthe␈αuser
␈↓ α←␈↓can␈α∞in␈α∞fact␈α∞do␈α∞this.␈α∞ The␈α∞criterion␈α∞may␈α∞appear␈α∞bizarre␈α∞and␈α∞may␈α∞in␈α∞fact␈α∞be
␈↓ α←␈↓achieving␈αsome␈αother␈αeffect␈αindirectly␈αvia␈αthat␈αvariable␈αname.␈α But␈αnote␈αthat
␈↓ α←␈↓the␈αcriterion␈αmust␈αbe␈αencoded␈αexplicitly␈αin␈αthe␈αpredicate␈αdefined␈αby␈αthe␈αuser,
␈↓ α←␈↓and␈αhence␈αthere␈αwill␈αat␈αleast␈αbe␈αan␈αexplicit␈αspecification␈αof␈αthe␈αcriterion␈α
used.
␈↓ α←␈↓This␈α
is␈αan␈α
improvement␈αover␈α
the␈α
situation␈αwhere␈α
a␈αlist␈α
is␈αhand-ordered,␈α
with
␈↓ α←␈↓no␈α
record␈α∞left␈α
behind.␈α
 We␈α∞have␈α
supplied␈α
a␈α∞means␈α
of␈α∞specifying␈α
invocation
␈↓ α←␈↓criteria␈α∩and␈α∩a␈α∩technique␈α⊃that␈α∩can,␈α∩if␈α∩used␈α⊃correctly,␈α∩insure␈α∩that␈α∩KSs␈α⊃are
␈↓ α←␈↓accurately␈α∩described␈α∩(i.e.,␈α∩by␈α∩referencing␈α∩the␈α∩code␈α∩directly).␈α∩There␈α∩are␈α⊃no
␈↓ α←␈↓constraints␈α⊂on␈α⊂how␈α⊂the␈α∂user␈α⊂may␈α⊂choose␈α⊂to␈α∂apply␈α⊂that␈α⊂capability,␈α⊂but␈α∂any
␈↓ α←␈↓deviousness will at least be more explicit.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∃further␈α∃objection␈α∃to␈α∀our␈α∃approach␈α∃might␈α∃claim␈α∃that␈α∀allowing
␈↓ α←␈↓arbitrary␈αpredicates␈α
to␈αexamine␈α
and␈αcharacterize␈αa␈α
KS␈αmay␈α
result␈αin␈α
odd␈αor
␈↓ α←␈↓unanticipated␈α∪characterizations.␈α∪ This␈α∪may␈α∩in␈α∪fact␈α∪be␈α∪an␈α∪advantage␈α∩and
␈↓ α←␈↓raises␈α
an␈αinteresting␈α
point.␈α
 The␈αuse␈α
of␈α
knowledge␈αin␈α
odd␈α
or␈αunexpected␈α
ways
␈↓ α←␈↓is␈α∞one␈α
aspect␈α∞of␈α
that␈α∞elusive␈α∞quality,␈α
creativity.␈α∞ With␈α
this␈α∞scheme,␈α∞we␈α
have
␈↓ α←␈↓taken␈α
one␈α
small␈α
step␈α
toward␈α
making␈αit␈α
possible␈α
for␈α
the␈α
system␈α
to␈αdiscover␈α
that
␈↓ α←␈↓a␈α∩particular␈α∩KS␈α∩has␈α∩a␈α∩characteristic␈α∩that␈α∩may␈α∩not␈α∩have␈α∩occurred␈α∩to␈α∩the
␈↓ α←␈↓programmer␈α∩who␈α∩wrote␈α∩the␈α∩KS␈α∩code.␈α∩ As␈α∩a␈α∩result,␈α∩the␈α∩system␈α∩may␈α⊃find
␈↓ α←␈↓unexpected applications for its knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?Even␈αwith␈αthe␈αbest␈αof␈αintentions,␈αhowever,␈αthere␈αremains␈αthe␈αquestion
␈↓ α←␈↓of␈α
the␈α
correctness␈αof␈α
the␈α
code␈α
that␈αimplements␈α
the␈α
content␈αexamination.␈α
 That
␈↓ α←␈↓is,␈αwhat␈αthe␈αprogrammer␈αintends␈αto␈αsay␈αin␈αthis␈αinvocation␈αcriteria␈α``language''
␈↓ α←␈↓and␈α∞what␈α
he␈α∞actually␈α
writes␈α∞may␈α
be␈α∞slightly␈α
different.␈α∞ Have␈α
we,␈α∞then,␈α
only
␈↓ α←␈↓pushed␈α
the␈α
problem␈α
back␈α
one␈α
level?␈α
 No,␈α
because␈α
it␈α
has␈α
been␈α
formalized␈α
as
␈↓ α←␈↓well.␈α
 Note␈α
that␈α
the␈α
``language''␈α
supplies␈α
a␈α
precise␈α
and␈α
explicit␈α∞expression␈α
of
␈↓ α←␈↓the␈α∞invocation␈α∂criteria.␈α∞ Correctness␈α∂is␈α∞at␈α∞least␈α∂a␈α∞well-specified␈α∂and␈α∞testable
␈↓ α←␈↓question.

␈↓"β␈↓ α←␈↓␈↓αExpressiveness and generalized invocation criteria␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∞of␈α
the␈α∞weaker␈α
claims␈α∞made␈α
above␈α∞concerned␈α∞the␈α
expressiveness
␈↓ α←␈↓of␈α∃content-directed␈α∃invocation␈α∃(and,␈α∀by␈α∃extension,␈α∃the␈α∃expressiveness␈α∀of
␈↓ α←␈↓generalized␈αinvocation␈αcriteria).␈α
 The␈αweakness␈αlies␈α
in␈αour␈αsuggesting␈αthat␈α
the
␈↓ α←␈↓programmer␈α
use␈αthe␈α
programming␈αlanguage␈α
itself␈αto␈α
express␈αretrieval␈α
criteria,
␈↓ α←␈↓without␈α∃having␈α∃supplied␈α∀any␈α∃guidance␈α∃or␈α∃hints␈α∀about␈α∃how␈α∃to␈α∃do␈α∀this
␈↓ α←␈↓effectively.
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    235␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α
combination␈α
of␈α
the␈α
meta-rules␈α
and␈α
␈↓¬LISP␈↓␈α
is,␈α
at␈α
best,␈α
adequate␈α
(in
␈↓ α←␈↓the␈α
sense␈α
of␈αbeing␈α
able␈α
to␈α
express␈αany␈α
computable␈α
predicate)␈α
and␈αextensible
␈↓ α←␈↓only␈αbecause,␈αfor␈αexample,␈αnew␈α
predicate␈αfunctions␈αcan␈αbe␈αadded␈αto␈α
the␈αrule
␈↓ α←␈↓language␈αby␈αwriting␈αthem␈αin␈α␈↓¬LISP␈↓.␈α It␈αwould␈αbe␈αbetter,␈αof␈αcourse,␈αif␈αthe␈αmeta-
␈↓ α←␈↓rule␈α↔language␈α↔already␈α↔included␈α↔a␈α↔set␈α↔of␈α↔well-designed␈α_primitives␈α↔that
␈↓ α←␈↓provided␈α⊃a␈α⊃good␈α⊃foundation␈α⊃for␈α⊃expressing␈α⊃invocation␈α⊃criteria.␈α⊃ Then␈α⊂we
␈↓ α←␈↓could␈α∪suggest␈α∀not␈α∪only␈α∪that␈α∀``it's␈α∪useful␈α∀to␈α∪be␈α∪able␈α∀to␈α∪define␈α∀your␈α∪own,
␈↓ α←␈↓generalized␈α∀invocation␈α∀criteria''␈α∀but␈α∀might␈α∀also␈α∀say,␈α∀``and␈α∀here's␈α∃a␈α∀well-
␈↓ α←␈↓designed␈αlanguage␈αthat␈αwill␈αhelp␈α
get␈αyou␈αstarted,␈αwithout␈αrestricting␈αyou␈α
since
␈↓ α←␈↓it's␈αalso␈αextensible.''  ␈α
It␈αwould␈αbe␈α
useful␈αto␈αbe␈α
able␈αto␈αoffer␈α
the␈αuser␈αan␈α
initial
␈↓ α←␈↓set␈α
of␈α
descriptors␈α(like␈α
goals,␈α
side␈α
effects,␈αspeed␈α
and␈α
space␈α
requirements,␈αetc.)
␈↓ α←␈↓that␈α
was␈α∞rich␈α
enough␈α
to␈α∞allow␈α
him␈α
to␈α∞express␈α
interesting␈α
criteria␈α∞easily␈α
(i.e.,
␈↓ α←␈↓without␈αhaving␈αto␈α
discover␈αand␈αimplement␈α
that␈αset␈αof␈α
primitives␈αon␈αhis␈α
own).
␈↓ α←␈↓We have not as yet developed such a set; this is a focus for continued work.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αproblem␈αis␈αdifficult␈αfor␈αmany␈αreasons,␈αincluding␈αthe␈αfact␈αthat␈αthe
␈↓ α←␈↓task␈α_of␈α_assembling␈α_the␈α_``vocabulary''␈α_of␈α_descriptors␈α_may␈α_appear␈α_quite
␈↓ α←␈↓imposing,␈α∞and␈α
perhaps␈α∞endless.␈α
 What␈α∞guarantee␈α
have␈α∞we␈α
that␈α∞there␈α∞are␈α
in
␈↓ α←␈↓fact␈α∀a␈α∀finite␈α∀number␈α∀of␈α∀descriptors␈α∀(i.e.,␈α∀a␈α∀finite␈α∀number␈α∃of␈α∀conceptual
␈↓ α←␈↓primitives␈α∪for␈α∪describing␈α∩knowledge),␈α∪rather␈α∪than␈α∩an␈α∪infinite␈α∪number␈α∩of
␈↓ α←␈↓special␈α
characteristics?  ␈αThere␈α
may␈αbe␈α
no␈αguarantee,␈α
but␈αthe␈α
benefits␈α
are␈αin
␈↓ α←␈↓any␈αcase␈αincremental--it␈αis␈αuseful␈αto␈αwrite␈αany␈αof␈αthe␈αsystem␈αstrategies␈αin␈αthis
␈↓ α←␈↓form, so even a subset of the entire collection is useful.

␈↓"β␈↓ α←␈↓␈↓αFlexibility␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?We␈αhave␈αemphasized␈αthe␈αbenefits␈αof␈αflexibility,␈αbut␈αit␈αis␈αimportant␈αto
␈↓ α←␈↓consider␈αthese␈αbenefits␈αin␈αthe␈αappropriate␈αcontext.␈α While␈αit␈αmay␈αbe␈αobvious,
␈↓ α←␈↓it␈α
is␈αworth␈α
noting␈α
that␈αthe␈α
task␈α
at␈αhand␈α
must␈α
somehow␈αrequire␈α
this␈α
kind␈αof
␈↓ α←␈↓flexibility.␈α
 In␈α
our␈α∞particular␈α
application␈α
it␈α
arises␈α∞from␈α
the␈α
emphasis␈α∞on␈α
the
␈↓ α←␈↓necessity␈α⊂of␈α⊂incremental␈α⊂construction␈α⊂of␈α⊂large␈α⊂performance␈α⊂programs.␈α⊂ The
␈↓ α←␈↓central␈α
point␈αis␈α
that␈α
there␈αwill␈α
be␈αa␈α
large␈α
number␈αof␈α
knowledge␈α
sources␈αand
␈↓ α←␈↓strategies,␈α
with␈α
frequent␈α
changes␈α
to␈αboth␈α
over␈α
an␈α
extended␈α
period␈α
of␈αtime.␈α
 In
␈↓ α←␈↓this␈α∩case␈α∩there␈α∩is␈α∩a␈α∩premium␈α∩on␈α∩the␈α∩ability␈α∩to␈α∩organize␈α∩and␈α⊃manipulate
␈↓ α←␈↓knowledge.␈α
 This␈α
is␈α
not␈α
necessarily␈α
true␈α
for␈α
smaller,␈α
well-established␈αsystems
␈↓ α←␈↓whose knowledge base has stabilized.

␈↓"β␈↓ α←␈↓␈↓αSpeed␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?Speed␈αis␈αa␈αlimiting␈αfactor␈αhere␈αbecause␈αall␈αof␈αthe␈α
advantages␈αclaimed
␈↓ α←␈↓for␈αcontent-directed␈αinvocation␈α
depend␈αon␈αthe␈α
indirection␈αinherent␈αin␈α
content
␈↓ α←␈↓reference.␈α⊃ Examining␈α⊃KS␈α⊃code␈α⊃to␈α⊂deduce␈α⊃some␈α⊃subtle␈α⊃property␈α⊃can␈α⊃be␈α⊂a
␈↓ α←␈↓time-consuming␈α
affair,␈α
especially␈α
when␈α
compared␈α
to␈α
the␈α
speed␈α
of␈αreference␈α
by
␈↓ α←␈↓name.
␈↓"β␈↓ α←␈↓␈↓ β?However,␈αlike␈αmany␈αsimilar␈αconstructs,␈αmost␈αof␈αthe␈αcomputational␈αcost
␈↓ α←␈↓can␈α⊗be␈α⊗paid␈α⊗in␈α∃a␈α⊗background␈α⊗computation␈α⊗between␈α⊗performance␈α∃runs.
␈↓ α←␈↓During␈α
that␈α
time␈α
the␈αsystem␈α
could␈α
compute␈α
the␈αsets␈α
of␈α
KSs␈α
determined␈αby␈α
the
␈↓ α←␈↓␈↓236    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓various␈α
descriptions.␈↓
22␈↓␈αThe␈α
results␈α
might␈αeither␈α
be␈αsaved␈α
for␈α
execution␈αtime
␈↓ α←␈↓use,␈α⊂or,␈α∂in␈α⊂a␈α∂form␈α⊂of␈α∂``pre-compiling,''␈α⊂the␈α∂source␈α⊂code␈α∂might␈α⊂be␈α∂rewritten,
␈↓ α←␈↓replacing␈α⊃the␈α⊃descriptions␈α⊂with␈α⊃the␈α⊃sets␈α⊃they␈α⊂define.␈α⊃This␈α⊃offers␈α⊃both␈α⊂the
␈↓ α←␈↓flexibility of reference by description and the speed of reference by name.
␈↓"β␈↓ α←␈↓␈↓ β?Other␈αconstructs␈α(e.g.,␈α␈↓¬LISP␈↓␈αrecord␈αstructures)␈αoffer␈αsimilar␈αadvantages,
␈↓ α←␈↓in␈α∪providing␈α∀a␈α∪level␈α∀of␈α∪insulation␈α∀from␈α∪the␈α∀effects␈α∪of␈α∀changes,␈α∪without
␈↓ α←␈↓imposing␈α
execution␈αtime␈α
overhead.␈α
 This␈αparticular␈α
example,␈αhowever,␈α
effects
␈↓ α←␈↓a␈α∞very␈α∞standard␈α∞compiler-style␈α
transformation:␈α∞replacing␈α∞a␈α∞constant␈α∞with␈α
its
␈↓ α←␈↓value.␈α
 Since␈α
(by␈α
our␈αdefinitions)␈α
the␈α
total␈α
KS␈α
set␈αis␈α
fixed␈α
at␈α
compile␈αtime,␈α
the
␈↓ α←␈↓descriptions␈α∞in␈α∞a␈α∞strategy␈α∞are␈α∞really␈α∞constants␈α∞and␈α∞can␈α∞be␈α∞replaced␈α∂by␈α∞their
␈↓ α←␈↓values.␈α≤ This␈α≤offers␈α≠the␈α≤symbolic␈α≤analogue␈α≠of␈α≤the␈α≤ability␈α≤to␈α≠write
␈↓ α←␈↓(SQRT(PI)/7),␈α→with␈α→the␈α→compiler␈α→doing␈α→the␈α→work␈α→to␈α→replace␈α→it␈α→with
␈↓ α←␈↓.253207683.  The benefits of flexibility and clarity are identical.
␈↓"β␈↓ α←␈↓␈↓ β?It␈α∞may␈α∞be␈α∞useful,␈α
then,␈α∞to␈α∞picture␈α∞this␈α
process␈α∞as␈α∞a␈α∞generalization␈α
of
␈↓ α←␈↓the␈α∀usual␈α∀view␈α∀that␈α∀there␈α∀are␈α∀two␈α∀levels␈α∀of␈α∀code␈α∀(source␈α∀and␈α∀machine
␈↓ α←␈↓language)␈α
and␈α
work␈α
with␈α
multiple␈α
source␈α
codes,␈α
each␈α
with␈α
its␈α
own␈αcompiler.
␈↓ α←␈↓This is indicated in Fig. 7-5, with appropriate numbers to indicate level.


␈↓"␈↓ α←␈↓∧...source␈↓
ii␈↓∧
␈↓"␈↓ α←␈↓∧        ==compiler␈↓
ii␈↓∧==@
␈↓"␈↓ α←␈↓∧                   source␈↓
i␈↓∧
␈↓"␈↓ α←␈↓∧                        ==compiler␈↓
i␈↓∧==@
␈↓"␈↓ α←␈↓∧                                   source
␈↓"␈↓ α←␈↓∧                                        ==compiler==@
␈↓"␈↓ α←␈↓∧                                               machine language


␈↓"␈↓ α←␈↓α␈↓ ¬_Fig. 7-5.    Levels of code.    

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[22]␈α∂There␈α∂are␈α∂nontrivial␈α∂problems␈α∂associated␈α∂with␈α∂the␈α∂appearance␈α∂of␈α∞free
␈↓ α←␈↓variables in the description parts, as in: 
␈↓"β␈↓ α←␈↓␈↓↓if␈αconditions␈αA␈α
and␈αB␈αhold,␈αuse␈α
any␈αKS␈αwhich␈αdeals␈α
with␈αwhite␈αcell␈α
counts␈αat
␈↓ α←␈↓↓least as high as that of the current patient.␈↓
␈↓"β␈↓ α←␈↓It␈α⊂is␈α⊂still␈α⊂possible,␈α⊂however,␈α⊂to␈α⊂turn␈α⊂this␈α⊂into␈α⊂a␈α⊂list␈α⊂of␈α⊂specific␈α⊃KS␈α⊂names.
␈↓ α←␈↓Consider␈α⊂the␈α⊂total␈α⊂range␈α⊂of␈α⊂the␈α∂white␈α⊂cell␈α⊂count.␈α⊂ Each␈α⊂KS␈α⊂in␈α⊂the␈α∂system
␈↓ α←␈↓which␈α∞deals␈α∞with␈α∞it␈α∞will␈α∞be␈α∞relevant␈α∞to␈α∞a␈α∞specific␈α∞value␈α∞or␈α∞range␈α∞of␈α∞values.
␈↓ α←␈↓Thus,␈α
we␈α
can␈αdivide␈α
up␈α
the␈αtotal␈α
range␈α
into,␈αsay,␈α
m␈α
different␈α
segments,␈αand
␈↓ α←␈↓in␈αeach␈α
segment␈αa␈α
specified␈αset␈α
of␈αKSs␈α
will␈αbe␈α
relevant.␈α The␈α
single␈αstrategy
␈↓ α←␈↓above would have to be replaced with m strategies of the form,
␈↓"β␈↓ α←␈↓␈↓↓if␈α⊂conditions␈α∂A␈α⊂and␈α⊂B␈α∂hold,␈α⊂and␈α∂the␈α⊂white␈α⊂cell␈α∂count␈α⊂is␈α∂in␈α⊂range␈↓i␈↓↓␈α⊂then␈α∂use
␈↓ α←␈↓↓{KS␈↓i1␈↓↓, KS␈↓i2␈↓↓,...}
␈↓"β␈↓ α←␈↓where␈αrange␈↓i␈↓␈αis␈αa␈α
segment␈αof␈αthe␈αtotal␈α
range␈αand␈αthe␈αdescriptions␈α
have␈αbeen
␈↓ α←␈↓replaced␈α
with␈α∞the␈α
sets␈α
of␈α∞relevant␈α
KSs.␈α
 The␈α∞same␈α
technique␈α
works␈α∞for␈α
any
␈↓ α←␈↓continuous␈α
variable,␈α
of␈α
course;␈αdiscrete␈α
valued␈α
variables␈α
are␈α
a␈αdegenerate␈α
case
␈↓ α←␈↓in␈αwhich␈αall␈αthe␈αKSs␈αare␈αrelevant␈αto␈αone␈αor␈αmore␈αsingle␈αvalues␈αrather␈αthan␈αa
␈↓ α←␈↓range.
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    237␈↓

␈↓"β␈↓ α←␈↓Each␈αsource␈α
further␈αto␈α
the␈αleft␈α
is␈αtypically␈α
easier␈αto␈α
read␈αand␈αunderstand,␈α
and
␈↓ α←␈↓more␈αflexible,␈αbut␈αalso␈αslower.␈αIt␈αseems␈αplausible␈αto␈αsuggest␈αthat␈αmany␈αof␈αthe
␈↓ α←␈↓techniques␈α≡accepted␈α≡for␈α≡compilation␈α≡and␈α≡optimization␈α≡of␈α≡arithmetic
␈↓ α←␈↓expressions␈α∞may␈α∞have␈α∞analogues␈α∞in␈α∞symbolic␈α∞computation␈α∞as␈α∂well.␈α∞ (Related
␈↓ α←␈↓ideas are found in [Low74] and [Samet75].)
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂generalized␈α⊂view␈α∂of␈α∂Fig.␈α⊂7-5␈α∂also␈α∂demonstrates␈α⊂another␈α∂point: 
␈↓ α←␈↓The␈α
distinction␈αwe␈α
have␈α
drawn␈αbetween␈α
compile␈α
time␈αand␈α
execution␈α
time␈αis
␈↓ α←␈↓one␈αthe␈α
user␈αcan␈α
determine␈αon␈α
his␈αown.␈α
The␈αdecision␈α
of␈αwhat␈α
level␈αof␈αcode␈α
to
␈↓ α←␈↓execute␈α⊃is␈α∩entirely␈α⊃open--the␈α∩higher␈α⊃level␈α⊃versions␈α∩are␈α⊃more␈α∩flexible,␈α⊃the
␈↓ α←␈↓lower␈α∂level␈α∂versions␈α∂are␈α∂faster.␈α∂There␈α∞is␈α∂currently␈α∂no␈α∂way␈α∂to␈α∂have␈α∂both␈α∞at
␈↓ α←␈↓once, but given good compilers, the transformation is relatively painless.␈↓
23␈↓
␈↓"β␈↓ α←␈↓␈↓ β?In␈α∃general,␈α∀then,␈α∃the␈α∀dividing␈α∃line␈α∀can␈α∃be␈α∀drawn␈α∃at␈α∃any␈α∀point.
␈↓ α←␈↓Whatever␈αis␈αrelegated␈αto␈α
being␈αa␈αcompile␈αtime␈α
transformation␈αcan␈αbe␈αdone␈α
in
␈↓ α←␈↓background mode and hence costs nothing at execution time.
␈↓"β␈↓ α←␈↓␈↓ β?Note␈α∂that␈α∂this␈α∂also␈α∂has␈α∂some␈α∞impact␈α∂on␈α∂the␈α∂issue␈α∂of␈α∂complexity.␈α∞ It
␈↓ α←␈↓uses␈α⊂a␈α⊂succession␈α⊃of␈α⊂high-level␈α⊂languages,␈α⊃each␈α⊂of␈α⊂which␈α⊃suppresses␈α⊂more
␈↓ α←␈↓implementation␈α
detail␈α
and␈α
concentrates␈α∞instead␈α
on␈α
the␈α
more␈α∞important␈α
issue
␈↓ α←␈↓of knowledge organization and use.

␈↓"β␈↓ α←␈↓␈↓αLimitations:  Summary␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂point␈α∂then␈α∂is␈α∂not␈α∂that␈α∂content-directed␈α∂invocation␈α∂ought␈α⊂to␈α∂be
␈↓ α←␈↓used␈αexclusively,␈αnor␈αthat␈αit␈αreplace␈αconcepts␈αlike␈αgoals,␈αpatterns,␈αetc.␈α We␈αare
␈↓ α←␈↓concerned␈α
with␈α∞the␈α
validity␈α∞of␈α
the␈α∞connection␈α
between␈α∞KS␈α
handles␈α∞and␈α
KS
␈↓ α←␈↓code,␈α∪how␈α∀expressive␈α∪the␈α∪handles␈α∀are,␈α∪and␈α∪what␈α∀they␈α∪contribute␈α∀to␈α∪the
␈↓ α←␈↓flexibility␈α
of␈α
the␈αsystem␈α
in␈α
the␈αface␈α
of␈α
changes␈αto␈α
the␈α
knowledge␈αbase.␈α
 Where
␈↓ α←␈↓possible,␈αthen,␈α
the␈αapproach␈αof␈α
content-directed␈αinvocation␈α
provides␈αa␈αway␈α
of
␈↓ α←␈↓assuring␈α∀a␈α∀degree␈α∪of␈α∀validity,␈α∀can␈α∀offer␈α∪a␈α∀more␈α∀expressive␈α∀syntax,␈α∪and
␈↓ α←␈↓provides a flexibility that can be very useful.

␈↓"β␈↓ α←␈↓␈↓α7-5-5    Future applications:  Choosing control regimes␈↓
␈↓"β␈↓ α←␈↓␈↓ β?We␈α∞described␈α∞the␈α∞use␈α∞of␈α∞strategies␈α∞as␈α∞a␈α∞means␈α∞of␈α∞``tuning''␈α∞a␈α
control
␈↓ α←␈↓structure␈αand␈αsuggested␈αin␈αaddition␈αthat␈αcontent␈αreference␈αoffered␈αa␈αmeans␈α
of
␈↓ α←␈↓defining␈α∃generalized␈α∃invocation␈α∃criteria.␈α∃ While␈α∃it␈α∃has␈α∃not␈α∃as␈α∃yet␈α∀been
␈↓ α←␈↓implemented,␈α∩we␈α∩can␈α∩imagine␈α∩pushing␈α⊃this␈α∩one␈α∩step␈α∩further.␈α∩ By␈α⊃adding
␈↓ α←␈↓another␈α
layer␈αof␈α
rules␈αabove␈α
the␈α
invocation␈αcriteria,␈α
we␈αmight␈α
gain␈αthe␈α
ability
␈↓ α←␈↓to␈αchoose␈αfrom␈αamong␈αthe␈αset␈αof␈αcriteria,␈αthat␈αis,␈αto␈αchoose␈αthe␈αcontrol␈αregime
␈↓ α←␈↓to␈αuse.␈α There␈αwould␈αbe␈αa␈αnumber␈αof␈αrules␈αdefining␈αvarious␈αretrieval␈αcriteria
␈↓ α←␈↓(goal,␈α∞event,␈α
etc.)␈α∞and␈α∞a␈α
number␈α∞of␈α∞(meta-)rules␈α
which␈α∞selected␈α∞from␈α
among
␈↓ α←␈↓these␈αrules␈αand␈αhence␈αchose␈αa␈αcontrol␈αstructure.␈α This␈αwould␈αmake␈αit␈αpossible

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[23]␈α∃Work␈α∀in␈α∃this␈α∃vein␈α∀is␈α∃described␈α∃in␈α∀[Mitchell70],␈α∃which␈α∃explores␈α∀a
␈↓ α←␈↓programming␈α
language␈α
that␈α
makes␈α
possible␈α
the␈α
execution␈α
of␈α
different␈α
levels
␈↓ α←␈↓of␈α∂code␈α∞in␈α∂a␈α∞manner␈α∂that␈α∞is␈α∂transparent␈α∞to␈α∂the␈α∞user.␈α∂ See␈α∂also␈α∞[Hansen74],
␈↓ α←␈↓which␈α∂describes␈α∂a␈α∂compiler␈α∞that␈α∂incrementally␈α∂compiles␈α∂and␈α∂optimizes␈α∞code
␈↓ α←␈↓according to the patterns of use and modification of the code.
␈↓ α←␈↓␈↓238    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓for␈α∀the␈α∪program␈α∀to␈α∪dynamically␈α∀change␈α∪control␈α∀structures␈α∪as␈α∀a␈α∪problem
␈↓ α←␈↓progressed␈α⊃(an␈α⊃ability␈α⊃lacking␈α∩in␈α⊃the␈α⊃current␈α⊃implementation,␈α⊃as␈α∩noted␈α⊃in
␈↓ α←␈↓Section 7-4-6).
␈↓"β␈↓ α←␈↓␈↓ β?Such␈α∂``execution␈α∂time␈α∂flexibility''␈α⊂would␈α∂make␈α∂possible␈α∂a␈α⊂number␈α∂of
␈↓ α←␈↓interesting␈α∃abilities.␈α⊗ Consider,␈α∃for␈α∃example,␈α⊗a␈α∃program␈α∃designed␈α⊗to␈α∃do
␈↓ α←␈↓heuristic␈α∞search.␈α
 For␈α∞domains␈α∞in␈α
which␈α∞a␈α∞single␈α
search␈α∞procedure␈α∞does␈α
not
␈↓ α←␈↓provide␈αeffective␈αperformance,␈α
it␈αwould␈αprove␈αvery␈α
useful␈αfor␈αthe␈αprogram␈α
to
␈↓ α←␈↓be␈αable␈α
to␈αdecide␈αfrom␈α
moment␈αto␈α
moment␈αwhat␈αform␈α
of␈αsearch␈α
would␈αmost
␈↓ α←␈↓likely␈α
be␈α
successful.␈α
 It␈α∞might␈α
thus␈α
use␈α
branch␈α∞and␈α
bound␈α
at␈α
one␈α∞point,␈α
hill
␈↓ α←␈↓climbing␈α∂at␈α∂another,␈α∂and␈α∂so␈α∂on.␈α∂ Note␈α∞that␈α∂we␈α∂are␈α∂not␈α∂speaking␈α∂of␈α∂a␈α∞pre-
␈↓ α←␈↓programmed␈α∀succession␈α∀of␈α∀techniques,␈α∀but␈α∀an␈α∀ability␈α∀to␈α∀choose␈α∀each␈α∀in
␈↓ α←␈↓response to changes in the problem state.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∃more␈α∃relevant␈α∀example␈α∃would␈α∃be␈α∀a␈α∃system␈α∃that␈α∃attempted␈α∀to
␈↓ α←␈↓simulate␈αthe␈α
diagnostic␈αbehavior␈α
of␈αan␈α
experienced␈αclinician.␈α
Such␈αbehavior
␈↓ α←␈↓has␈α∞been␈α∂found␈α∞([Rubin75],␈α∂[Miller75])␈α∞to␈α∂be␈α∞a␈α∂complex␈α∞shifting␈α∂back␈α∞and
␈↓ α←␈↓forth␈α≠between␈α≠processes␈α≠of␈α≠collecting␈α≠data␈α≠(event-driven␈α≠implication),
␈↓ α←␈↓establishing␈α
causal␈α
chains␈α
(heuristic␈α
search),␈α
verifying␈α
(goal-directed␈αsearch),
␈↓ α←␈↓and␈α∩others.␈α∩This␈α∩is␈α∩reminiscent␈α∩of␈α∩the␈α∩general␈α∩perspective␈α∩on␈α⊃knowledge
␈↓ α←␈↓organization␈α∞described␈α∂earlier,␈α∞since␈α∞the␈α∂physician␈α∞is␈α∞quite␈α∂clearly␈α∞changing
␈↓ α←␈↓his␈αentire␈αapproach␈αunder␈αthe␈αcontrol␈αof␈αstrategies␈α(``clinical␈αexperience'')␈αthat
␈↓ α←␈↓choose the most appropriate response.
␈↓"β␈↓ α←␈↓␈↓ β?Few␈α⊂computational␈α⊂systems␈α∂have␈α⊂offered␈α⊂very␈α∂much␈α⊂of␈α⊂this␈α⊂sort␈α∂of
␈↓ α←␈↓flexibility,␈α∩however.␈α⊃ Typically,␈α∩this␈α⊃results␈α∩from␈α⊃three␈α∩interrelated␈α⊃causes.
␈↓ α←␈↓The␈α↔problem␈α↔is␈α⊗that␈α↔often␈α↔an␈α↔extensive␈α⊗amount␈α↔of␈α↔knowledge␈α↔is␈α⊗(a)
␈↓ α←␈↓embedded␈α
in␈α
the␈α
control␈α
structure,␈α
(b)␈α
represented␈α
there␈α
only␈α
implicitly,␈α
and
␈↓ α←␈↓(c)␈α
treated␈αon␈α
a␈αspecial-case␈α
basis.␈αIn␈α
fact,␈αmuch␈α
of␈αwhat␈α
is␈αtypically␈α
described
␈↓ α←␈↓as␈α∪clever␈α∪coding␈α∪or␈α∪implementation␈α∪technique␈α∪often␈α∪represents␈α∩important
␈↓ α←␈↓insights␈α→into␈α→useful␈α→strategies.␈α_ For␈α→example,␈α→[Green69],␈α→in␈α_describing
␈↓ α←␈↓modifications␈α↔to␈α↔resolution␈α↔to␈α↔allow␈α↔his␈α↔system␈α↔to␈α↔deal␈α↔with␈α↔program
␈↓ α←␈↓construction,␈α∨mentions␈α≡in␈α∨passing␈α≡several␈α∨interesting␈α≡domain-specific
␈↓ α←␈↓strategies.␈α Expressing␈αsuch␈αchunks␈αof␈αknowledge␈αas␈αdistinct␈αstrategies␈αhas␈αin
␈↓ α←␈↓the␈α∩past␈α∩been␈α∩inhibited␈α∩by␈α∪the␈α∩lack␈α∩of␈α∩a␈α∩convenient␈α∩formalism␈α∪and␈α∩the
␈↓ α←␈↓difficulty␈α∩of␈α⊃generalizing␈α∩what␈α∩may␈α⊃have␈α∩been␈α∩seen␈α⊃initially␈α∩as␈α∩simply␈α⊃a
␈↓ α←␈↓special-case coding hack.
␈↓"β␈↓ α←␈↓␈↓ β?Yet␈α
the␈α
desire␈α
for␈α
increased␈αflexibility␈α
has␈α
a␈α
long␈α
history.␈α
Much␈αof␈α
the
␈↓ α←␈↓early␈α
speculation␈αabout␈α
programs␈αthat␈α
could␈α
learn␈αcentered␈α
around␈αthe␈α
ability
␈↓ α←␈↓to␈αshift␈αstrategies␈αdynamically␈αas␈αthe␈αsituation␈αrequired.␈α Gelernter␈α
mentioned
␈↓ α←␈↓it␈αexplicitly␈α
in␈α1959␈α[Gelernter59];␈α
ten␈αyears␈α
later␈αGreen␈α([Green69])␈α
discussed
␈↓ α←␈↓somewhat␈α→more␈α→concrete␈α→proposals␈α→centered␈α→around␈α→the␈α→possibility␈α→of
␈↓ α←␈↓expressing␈α
strategies␈α
in␈αthe␈α
same␈α
first-order␈α
predicate␈αcalculus␈α
that␈α
the␈αrest␈α
of
␈↓ α←␈↓his␈α⊂system␈α⊃used.␈α⊂ Meta-rules␈α⊂of␈α⊃the␈α⊂sort␈α⊂described␈α⊃in␈α⊂Section␈α⊃7-4-4␈α⊂make
␈↓ α←␈↓possible␈αa␈αlimited␈αform␈αof␈αexecution␈αtime␈αflexibility.␈α They␈αmake␈αpart␈αof␈αthe
␈↓ α←␈↓system's␈α∞control␈α∞structure␈α∞explicit␈α∞and␈α∞accessible,␈α∞and␈α∞hence␈α∞manipulable␈α∞by
␈↓ α←␈↓other meta-rules.
␈↓"β␈↓ α←␈↓␈↓ β?What␈α
we␈α
are␈α
looking␈α
for,␈α
then,␈α
are␈α
``softwired''␈α
strategies.␈α Achieving
␈↓ α←␈↓␈↓7-5␈↓ π[BROADER IMPLICATIONS    239␈↓

␈↓"β␈↓ α←␈↓this␈αobjective␈αis␈αpredicated␈αon␈αovercoming␈αthe␈αthree␈αproblems␈αmentioned,␈αyet
␈↓ α←␈↓this␈α
is␈αoften␈α
not␈αdifficult␈α
to␈α
do.␈αFirst,␈α
the␈αknowledge␈α
should␈αbe␈α
isolated␈α
as␈αa
␈↓ α←␈↓distinct␈α
chunk,␈α
rather␈α
than␈α
embedded␈α
in␈αthe␈α
system␈α
code␈α
(e.g.,␈α
by␈α
putting␈αit␈α
in
␈↓ α←␈↓a␈α_single␈α_meta-rule).␈α↔ Second,␈α_the␈α_knowledge␈α↔should␈α_be␈α_made␈α↔explicit,
␈↓ α←␈↓identifying␈α∞exactly␈α∞what␈α∞it␈α
is␈α∞that␈α∞the␈α∞system␈α
should␈α∞know␈α∞(e.g.,␈α∞the␈α
criteria
␈↓ α←␈↓for deciding between control structures).
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α∀this␈α∀well-defined␈α∀chunk␈α∪should␈α∀be␈α∀generalized␈α∀as␈α∀far␈α∪as
␈↓ α←␈↓possible,␈αto␈αprovide␈αmaximal␈αpower.␈α By␈α``generalize,''␈αwe␈αmean␈αthat␈αit␈αshould
␈↓ α←␈↓be␈αconsidered␈αnot␈αas␈αa␈αspecial␈αcase␈αor␈αclever␈αtrick␈αbut,␈αas␈αfar␈αas␈αpossible,␈αas␈αa
␈↓ α←␈↓fundamental principle of knowledge organization.
␈↓"β␈↓ α←␈↓␈↓ β?If␈α∞all␈α∞of␈α∞this␈α∞is␈α∞really␈α∞so␈α∞useful,␈α∞why␈α∞isn't␈α∞it␈α∞common␈α∂practice?␈α∞ The
␈↓ α←␈↓primary␈α→reason␈α→is␈α→that␈α→new␈α→problem-solving␈α→methods␈α→are␈α→still␈α→being
␈↓ α←␈↓uncovered␈α
and␈α
explored--their␈α
power,␈αpotential,␈α
and␈α
applicability␈α
are␈α
yet␈αto
␈↓ α←␈↓be␈αtotally␈αdefined.␈α We␈αare␈αproposing␈αa␈αframework␈αfor␈αintegrating␈αa␈αnumber
␈↓ α←␈↓of␈α∞methods,␈α∞based␈α∞on␈α∞the␈α∞belief␈α∂that␈α∞there␈α∞are␈α∞now␈α∞a␈α∞sufficient␈α∂number␈α∞of
␈↓ α←␈↓techniques␈α
understood␈α
well␈α
enough␈αthat␈α
we␈α
can␈α
begin␈α
examining␈αthe␈α
problem
␈↓ α←␈↓of integrating many of them into a cohesive system.
␈↓"β␈↓ α←␈↓␈↓ β?But␈αwhy␈α
consider␈αthe␈α
whole␈αtool␈αset,␈α
with␈αall␈α
the␈αassociated␈αcost?␈α
 The
␈↓ α←␈↓answer␈α∩may␈α∩lie␈α∪in␈α∩the␈α∩nature␈α∩of␈α∪the␈α∩problems␈α∩currently␈α∪being␈α∩explored.
␈↓ α←␈↓Chapter␈α∂1␈α⊂noted␈α∂the␈α∂trend␈α⊂toward␈α∂real␈α∂world␈α⊂problems␈α∂of␈α⊂significant␈α∂size.
␈↓ α←␈↓Where␈α⊂a␈α⊂single␈α⊂methodology␈α⊂may␈α∂prove␈α⊂powerful␈α⊂enough␈α⊂to␈α⊂handle␈α∂well-
␈↓ α←␈↓isolated␈α
tasks␈α
in␈αselected␈α
domains,␈α
the␈α
broader␈αscope␈α
of␈α
real␈α
world␈αproblems
␈↓ α←␈↓appears␈αto␈αrequire␈αa␈αcorrespondingly␈αbroader␈αcollection␈αof␈αabilities.␈α Strategic
␈↓ α←␈↓knowledge may provide a facility for directing their use.
␈↓"β␈↓ α←␈↓␈↓ β?Thus,␈αthe␈α
primary␈αadvantage␈α
of␈αincreased␈α
execution␈αtime␈αflexibility␈α
is
␈↓ α←␈↓the␈αpotential␈αfor␈αbuilding␈αmore␈αflexible␈αprograms.␈α Systems␈αequipped␈α
with␈αa
␈↓ α←␈↓single␈α∞``hardwired''␈α∞strategy,␈α∞hence␈α∞a␈α∞single␈α∞approach␈α∞to␈α∞the␈α∞problem,␈α
would
␈↓ α←␈↓appear␈α∞to␈α∞suffer␈α∞the␈α∞same␈α∞shortcomings␈α∞as␈α∞the␈α∞general␈α∞problem␈α∂solver␈α∞type
␈↓ α←␈↓programs.␈α
The␈α
latter␈α
proved␈α
unable␈α
to␈α
deal␈α
with␈α
difficult␈α∞problems␈α
because
␈↓ α←␈↓of␈α↔their␈α↔single,␈α↔domain-independent␈α↔methodology,␈α↔which␈α↔could␈α_not␈α↔be
␈↓ α←␈↓modified␈α∪by␈α∪important␈α∩and␈α∪useful␈α∪domain-specific␈α∪knowledge.␈α∩ Similarly,
␈↓ α←␈↓current␈α⊂task-specific␈α∂programs,␈α⊂with␈α∂their␈α⊂domain-specific␈α⊂methods,␈α∂cannot
␈↓ α←␈↓dynamically shift the way they function and have an analogous limitation.

␈↓"β␈↓ α←␈↓␈↓α7-5-6    Review␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃central␈α⊃aim␈α∩of␈α⊃this␈α⊃section␈α⊃has␈α∩been␈α⊃to␈α⊃consider␈α∩the␈α⊃broader
␈↓ α←␈↓implications␈α⊃that␈α⊃follow␈α⊃from␈α⊃the␈α⊃techniques␈α⊃used␈α⊃in␈α⊃implementing␈α⊂meta-
␈↓ α←␈↓rules.␈α" The␈α"discussion␈α"focused␈α"on␈α"two␈α!techniques--content-directed
␈↓ α←␈↓invocation␈α≡and␈α≡generalized␈α≡invocation␈α≡criteria--and␈α≡considered␈α≥their
␈↓ α←␈↓implications␈αas␈αprogramming␈αtechniques␈αin␈αgeneral,␈αindependent␈αof␈αthe␈αissue
␈↓ α←␈↓of encoding strategy knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α
began␈α
by␈αexamining␈α
different␈α
approaches␈αto␈α
referring␈α
to␈αKSs␈α
and
␈↓ α←␈↓defined␈αthe␈αidea␈αof␈αcontent␈αreference.␈α The␈αuse␈αof␈αthis␈αform␈αof␈αreference␈αas␈αa
␈↓ α←␈↓basis␈αfor␈α
KS␈αretrieval␈α
and␈αinvocation␈α
was␈αtermed␈αcontent-directed␈α
invocation
␈↓ α←␈↓and␈α⊃was␈α⊃compared␈α⊃to␈α⊃previous␈α⊃approaches␈α⊃to␈α⊃invocation.␈α⊃ This␈α⊃historical
␈↓ α←␈↓␈↓240    STRATEGIES␈↓ 
#7-5␈↓

␈↓"β␈↓ α←␈↓overview␈α∞suggested␈α
that␈α∞content-directed␈α∞invocation␈α
offers␈α∞an␈α∞added␈α
degree
␈↓ α←␈↓of␈α⊂validity␈α⊂and␈α⊂expressiveness␈α⊂to␈α⊂invocation.␈α⊂ The␈α⊂technique␈α⊂also␈α⊂offers␈α∂a
␈↓ α←␈↓way␈αof␈αmaking␈α
the␈αcriteria␈αfor␈α
KS␈αretrieval␈αaccessible␈α
and,␈αhence,␈αmore␈α
easily
␈↓ α←␈↓modifiable,␈α∪rather␈α∪than␈α∪predetermined␈α∪and␈α∪hardwired␈α∪into␈α∪the␈α∪language
␈↓ α←␈↓interpreter,␈αas␈αis␈αthe␈α
case␈αfor␈αmost␈αprogramming␈αlanguages.␈α
 Content-directed
␈↓ α←␈↓invocation␈α∂was␈α∂also␈α∞seen␈α∂to␈α∂offer␈α∞a␈α∂number␈α∂of␈α∞advantages␈α∂in␈α∂terms␈α∂of␈α∞the
␈↓ α←␈↓flexibility␈α
of␈α∞the␈α
resulting␈α
system: ␈α∞It␈α
is␈α
easier␈α∞to␈α
introduce␈α
changes␈α∞into␈α
the
␈↓ α←␈↓knowledge base in systems using this technique.
␈↓"β␈↓ α←␈↓␈↓ β?Generalized␈α∪invocation␈α∪criteria␈α∪arose␈α∪from␈α∪making␈α∪the␈α∪invocation
␈↓ α←␈↓criteria␈αaccessible␈αand␈αfrom␈αspecifying␈αthem␈αexplicitly␈αand␈αfunctionally.␈α This
␈↓ α←␈↓technique␈α∪was␈α∪seen␈α∩to␈α∪offer␈α∪significant␈α∩advantages␈α∪in␈α∪terms␈α∪of␈α∩effecting
␈↓ α←␈↓control structures explicitly rather than via a range of indirect effects.
␈↓"β␈↓ α←␈↓␈↓ β?Content-directed␈α≤invocation␈α≤and␈α≤generalized␈α≤invocation␈α≠criteria
␈↓ α←␈↓appear␈α
to␈α
be␈α
programming␈α
techniques␈α
that␈α
offer␈α
potential␈α
increases␈α∞in␈α
both
␈↓ α←␈↓the␈αvalidity␈αand␈αexpressiveness␈αof␈αthe␈αinvocation␈αprocess,␈αthat␈αoffer␈αincreases
␈↓ α←␈↓in␈αthe␈αflexibility␈αof␈αthe␈αsystem␈αin␈αresponse␈αto␈αchanges␈αin␈αthe␈αknowledge␈αbase,
␈↓ α←␈↓and that permit explicit specification of invocation criteria.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α
also␈α
examined␈α
several␈αlimitations␈α
of␈α
the␈α
current␈αimplementations
␈↓ α←␈↓and␈αexplored␈αdifficulties␈αthat␈αhave␈αto␈αbe␈αovercome␈αbefore␈αthe␈αfull␈αbenefits␈α
of
␈↓ α←␈↓these techniques can be realized.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈αwe␈α
speculated␈αabout␈αthe␈α
possible␈αuse␈α
of␈αmeta-rules␈αin␈α
creating
␈↓ α←␈↓a␈αsystem␈αcapable␈α
of␈αadaptively␈αchanging␈αits␈α
control␈αstructure␈αas␈α
the␈αproblem
␈↓ α←␈↓solution proceeds.
␈↓ α←␈↓␈↓7-6␈↓ π↑A TAXONOMY, OF SORTS    241␈↓

␈↓"β␈↓ α←␈↓␈↓α7-6    A TAXONOMY, OF SORTS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?In␈α≡this␈α≡section␈α≡the␈α≥discussion␈α≡broadens␈α≡beyond␈α≡the␈α≥specific
␈↓ α←␈↓implementation␈α∂of␈α∂strategies␈α∂demonstrated␈α∂by␈α∂meta-rules.␈α∂ We␈α∂examine␈α∞the
␈↓ α←␈↓kinds␈αof␈αstrategies␈αthat␈αhave␈αbeen␈αused␈αin␈αa␈αvariety␈αof␈αsystems,␈αwith␈αthe␈αaim
␈↓ α←␈↓of␈α∞developing␈α
a␈α∞rough␈α
taxonomy␈α∞of␈α∞strategy␈α
types.␈α∞ This,␈α
in␈α∞turn,␈α∞will␈α
help
␈↓ α←␈↓provide␈α
a␈αbetter␈α
understanding␈αof␈α
the␈α
range␈αof␈α
possible␈αstrategy␈α
types␈αand␈α
set
␈↓ α←␈↓in perspective the work on meta-rules.
␈↓"β␈↓ α←␈↓␈↓ β?The dimensions of the taxonomy are:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?generality,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?degree of explicitness,
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?organization, and
␈↓"β␈↓ α←␈↓␈↓ ββ(d)␈↓ β?character.

␈↓ α←␈↓While␈α
the␈α
set␈α
is␈α
not␈α
necessarily␈α
comprehensive,␈α
these␈α
four␈α
appear␈α
to␈αcapture␈α
a
␈↓ α←␈↓number␈α↔of␈α↔interesting␈α↔distinctions␈α↔and␈α↔help␈α↔characterize␈α↔the␈α↔range␈α↔of
␈↓ α←␈↓possibilities.␈α
 Examples␈αof␈α
actual␈αsystems␈α
are␈αused␈α
to␈αillustrate␈α
sample␈αpoints
␈↓ α←␈↓in each dimension.␈↓
24␈↓

␈↓"β␈↓ α←␈↓␈↓α7-6-1    Generality␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩most␈α∪general␈α∩strategies␈α∩are␈α∪those␈α∩we␈α∩will␈α∪term␈α∩␈↓↓representation
␈↓ α←␈↓↓independent␈↓,␈α∂since␈α∞they␈α∂describe␈α∂problem-solving␈α∞approaches␈α∂which␈α∂can␈α∞be
␈↓ α←␈↓used␈αno␈αmatter␈αwhat␈αunderlying␈αrepresentation␈αis␈αchosen.␈αAs␈αan␈αexample,␈α
the
␈↓ α←␈↓general␈α∩idea␈α∩of␈α∩a␈α⊃goal-directed␈α∩(``working␈α∩backward'')␈α∩vs.␈α∩a␈α⊃data-directed
␈↓ α←␈↓(``working␈α∩forward'')␈α∩approach␈α∩is␈α∩relevant␈α∩over␈α∩a␈α∩wide␈α∩range␈α∩of␈α⊃problem
␈↓ α←␈↓organizations␈α⊃and␈α∩representations.␈α⊃ Polya␈α∩[Polya54]␈α⊃describes␈α∩several␈α⊃such
␈↓ α←␈↓techniques;␈α∀a␈α∀recent␈α∀book␈α∪by␈α∀Wickelgren␈α∀[Wickelgren74]␈α∀deals␈α∀with␈α∪the
␈↓ α←␈↓subject␈α
in␈αmore␈α
general␈α
terms.␈αAn␈α
interesting␈α
and␈αvery␈α
abstract␈α
approach␈αis
␈↓ α←␈↓taken␈α∞in␈α∞␈↓↓Strategy␈α∞Notebook␈↓␈α∞[Interaction72],␈α∞which␈α∞explores␈α∞six␈α∂categories␈α∞of
␈↓ α←␈↓strategies␈α
and␈α
offers␈α
analyses␈α
of␈αthe␈α
strengths␈α
and␈α
weaknesses␈α
of␈α
each.␈α(The
␈↓ α←␈↓last␈α⊂category␈α⊂is␈α⊂labeled␈α⊂``metaheuristics''␈α⊂and␈α⊂describes␈α⊂several␈α⊂very␈α∂general
␈↓ α←␈↓second-order strategies.)
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊂work␈α⊃in␈α⊂[Green69]␈α⊃is␈α⊂a␈α⊂good␈α⊃example␈α⊂of␈α⊃the␈α⊂use␈α⊃of␈α⊂␈↓↓domain-
␈↓ α←␈↓↓independent␈↓␈αtechniques.␈α That␈αsystem␈αused␈αfirst-order␈αpredicate␈αcalculus␈αand
␈↓ α←␈↓resolution␈α
theorem␈α
proving.␈α Strategies␈α
employed␈α
included␈α
unit␈αpreference,␈α
set
␈↓ α←␈↓of␈α∞support,␈α∞a␈α
bound␈α∞on␈α∞the␈α
number␈α∞of␈α∞levels␈α
explored,␈α∞and␈α∞the␈α∞deletion␈α
of
␈↓ α←␈↓subsumed␈α
clauses.␈α Since␈α
these␈αtechniques␈α
are␈αmeaningless␈α
outside␈αthe␈α
context
␈↓ α←␈↓of␈α⊃theorem␈α⊃proving,␈α⊃they␈α⊃are␈α⊂not␈α⊃independent␈α⊃of␈α⊃the␈α⊃representation.␈α⊂The
␈↓ α←␈↓domain␈αindependence␈αof␈αthe␈αunderlying␈αmethodology,␈αhowever,␈αindicates␈αthe
␈↓ α←␈↓applicability␈α⊗of␈α⊗the␈α⊗techniques␈α↔to␈α⊗the␈α⊗theorem-proving␈α⊗process␈α↔in␈α⊗any
␈↓ α←␈↓domain.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α
second␈α
example␈α
of␈αdomain␈α
independence␈α
is␈α
the␈α
concept␈αof␈α
planning

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[24]␈α∂Note␈α∞that␈α∂many␈α∂of␈α∞the␈α∂references␈α∞cited␈α∂used␈α∂a␈α∞range␈α∂of␈α∂techniques␈α∞so
␈↓ α←␈↓that at times only selected aspects of each may be relevant.
␈↓ α←␈↓␈↓242    STRATEGIES␈↓ 
#7-6␈↓

␈↓"β␈↓ α←␈↓in␈α
abstraction␈α
spaces,␈αas␈α
described␈α
in␈α
[Sacerdoti73].␈αThis␈α
system␈α
was␈αbased␈α
on
␈↓ α←␈↓the␈αmeans-ends␈αanalysis␈αmethodology␈αof␈α␈↓¬STRIPS␈↓␈αand␈αused␈αa␈α
powerful␈αstrategy
␈↓ α←␈↓of␈αplanning␈α
at␈αmultiple␈α
levels␈αof␈α
detail.␈α Since␈αthe␈α
basic␈αtask␈α
in␈α␈↓¬STRIPS␈↓␈α
is␈αthe
␈↓ α←␈↓choice␈αof␈αan␈αoperator␈αwhose␈α
application␈αmoves␈αthe␈αsystem␈αcloser␈αto␈α
the␈αgoal,
␈↓ α←␈↓the␈αuse␈α
of␈αrepeated␈α
planning␈αat␈αsuccessively␈α
lower␈αlevels␈α
of␈αdetail␈α
becomes␈αa
␈↓ α←␈↓way␈α
of␈α
directing␈α∞the␈α
choice␈α
of␈α
operators␈α∞and␈α
is␈α
thus␈α∞a␈α
domain-independent
␈↓ α←␈↓strategy.
␈↓"β␈↓ α←␈↓␈↓ β?By␈α⊃contrast,␈α⊃Gelernter's␈α⊃early␈α⊃work␈α⊃on␈α⊃a␈α⊃geometry␈α∩theorem␈α⊃prover
␈↓ α←␈↓[Gelernter59]␈α
has␈αsome␈α
interesting␈αexamples␈α
of␈α␈↓↓domain-specific␈α
␈↓strategies.␈α It
␈↓ α←␈↓takes␈αa␈αproblem␈αreduction␈αapproach,␈αsetting␈αup␈αan␈αexhaustive␈αlist␈αof␈αpossible
␈↓ α←␈↓subgoals␈α∃on␈α∀the␈α∃basis␈α∀of␈α∃available␈α∀axioms␈α∃and␈α∀theorems.␈α∃Subgoals␈α∀are
␈↓ α←␈↓ordered,␈α∀however,␈α∀on␈α∀the␈α∀basis␈α∀of␈α∀several␈α∀domain-specific␈α∃criteria: ␈α∀For
␈↓ α←␈↓example,␈α∞goals␈α∞involving␈α∞vertical␈α∞angles␈α
are␈α∞chosen␈α∞first␈α∞because␈α∞they␈α
often
␈↓ α←␈↓turn out to have one-step proofs.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α⊂as␈α∂illustrated␈α⊂earlier,␈α∂meta-rules␈α⊂have␈α∂been␈α⊂used␈α⊂to␈α∂express
␈↓ α←␈↓strategies␈α∩that␈α∩are␈α∪␈↓↓goal␈α∩specific␈↓.␈α∩The␈α∩␈↓	THUSE␈↓␈α∪construct␈α∩in␈α∩␈↓¬PLANNER␈↓␈α∪and␈α∩the
␈↓ α←␈↓GOALCLASS␈α⊂statement␈α⊂of␈α⊂␈↓¬QA4␈↓␈α∂are␈α⊂programming␈α⊂language␈α⊂constructs␈α∂that
␈↓ α←␈↓offer the facility for writing similar sorts of goal-specific strategies.

␈↓"β␈↓ α←␈↓␈↓α7-6-2    Degree of explicitness␈↓
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∂second␈α∂dimension␈α∞concerns␈α∂the␈α∂degree␈α∞of␈α∂explicitness␈α∂with␈α∞which
␈↓ α←␈↓strategies␈α
are␈α
represented.␈α
␈↓↓Explicit␈↓␈α
strategies␈α
are␈α
those␈α
that␈α
are␈α∞embodied␈α
in
␈↓ α←␈↓their␈αown␈αdistinct␈αconstructs␈αand␈αthat␈αcan␈αbe␈αidentified␈αas␈αseparate␈αentities␈α
in
␈↓ α←␈↓the␈αsystem.␈α
 ␈↓↓Implicit␈αstrategies␈↓␈α
come␈αin␈α
two␈αvarieties.␈α
 Those␈αthat␈α
are␈αbound
␈↓ α←␈↓up␈α
in␈α
some␈αother␈α
aspect␈α
of␈αthe␈α
system␈α
(typically␈αthe␈α
control␈α
structure)␈αwe␈α
refer
␈↓ α←␈↓to␈α
as␈α␈↓↓implementationally␈α
implicit␈↓.␈αIn␈α
this␈α
case␈αthe␈α
essential␈αidea␈α
of␈αthe␈α
strategy
␈↓ α←␈↓has␈αbeen␈αcoded␈αexplicitly␈αin␈α
the␈αsystem,␈αbut␈αthat␈αcode␈αis␈α
embedded␈α(perhaps
␈↓ α←␈↓quite␈α
subtly)␈αin␈α
other␈αconstructs.␈α
 ␈↓↓Conceptually␈α
implicit␈↓␈αstrategies␈α
are␈αthose␈α
for
␈↓ α←␈↓which␈α∪there␈α∩is␈α∪no␈α∩distinct␈α∪encoding␈α∩mechanism␈α∪anywhere␈α∩in␈α∪the␈α∩system.
␈↓ α←␈↓Typically,␈αtheir␈αeffect␈αis␈αrealized␈αvia␈αsome␈αside-effect␈αof␈αa␈αprocess␈α
that␈αother
␈↓ α←␈↓parts␈α
of␈α
the␈α
system␈α
are␈α
tuned␈α
to␈αdetect.␈α
 Since␈α
one␈α
of␈α
the␈α
global␈α
themes␈αin␈α
this
␈↓ α←␈↓work␈α∞concerns␈α
the␈α∞virtues␈α
of␈α∞making␈α
all␈α∞knowledge␈α
in␈α∞a␈α
system␈α∞explicit,␈α
we
␈↓ α←␈↓will␈αmake␈αthe␈αcase␈αbelow␈αthat␈αthe␈αfirst␈αof␈αthese␈αthree␈αapproaches␈αis␈αthe␈αmost
␈↓ α←␈↓advantageous and that the last presents problems.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αwork␈αreported␈αin␈α[Howe73]␈α
is␈αa␈αgood␈αexample␈αof␈α
a␈αconceptually
␈↓ α←␈↓implicit␈αstrategy.␈α
 That␈αsystem␈α
used␈αa␈αproduction␈α
rule␈αstyle␈α
representation␈αof
␈↓ α←␈↓chemical␈α⊂reactions␈α⊂to␈α⊂solve␈α⊂chemical␈α⊂synthesis␈α⊂problems.␈α⊃ Certain␈α⊂reactions
␈↓ α←␈↓(which␈α
were␈α∞capable␈α
of␈α∞supplying␈α
major␈α
structural␈α∞features)␈α
were␈α∞chosen␈α
as
␈↓ α←␈↓sufficiently␈α⊂desirable␈α⊂that␈α⊃they␈α⊂were␈α⊂considered␈α⊂important␈α⊃to␈α⊂use,␈α⊂if␈α⊃at␈α⊂all
␈↓ α←␈↓possible.␈α This␈αstrategy␈αwas␈αeffected␈αby␈αleaving␈αpart␈αof␈αthe␈αprecondition␈αside
␈↓ α←␈↓of␈α≠these␈α≠rules␈α≠incompletely␈α≠specified.␈α≠ That␈α≠is,␈α≠the␈α≠``details''␈α≠of␈α~the
␈↓ α←␈↓preconditions␈αwere␈αsuppressed,␈αa␈αtechnique␈αthat␈αcan␈αbe␈αseen␈αas␈αa␈αspecial-case
␈↓ α←␈↓implementation␈α∪of␈α∪Sacerdoti's␈α∪levels␈α∪of␈α∪abstraction␈α∪idea.␈α∪ The␈α∪rules␈α∩thus
␈↓ α←␈↓appeared␈α∞applicable␈α
in␈α∞contexts␈α∞that␈α
were,␈α∞in␈α∞fact,␈α
not␈α∞exactly␈α∞correct.␈α
 The
␈↓ α←␈↓remainder␈α∀of␈α∀the␈α∃rule␈α∀was␈α∀tailored␈α∀to␈α∃check␈α∀for␈α∀these␈α∃mismatches␈α∀and
␈↓ α←␈↓␈↓7-6␈↓ π↑A TAXONOMY, OF SORTS    243␈↓

␈↓"β␈↓ α←␈↓attempted␈α∀to␈α∀take␈α∀care␈α∀of␈α∀all␈α∪the␈α∀necessary␈α∀details␈α∀before␈α∀the␈α∀rule␈α∪was
␈↓ α←␈↓invoked.␈α Since␈αthe␈αrules␈αwere␈αhand-coded␈αto␈αprovide␈αfor␈αthis␈αsuppression␈α
of
␈↓ α←␈↓detail,␈α∂the␈α∂strategy␈α⊂is␈α∂effected␈α∂indirectly␈α⊂and␈α∂has␈α∂no␈α⊂explicit␈α∂representation
␈↓ α←␈↓anywhere in the system.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α≡work␈α≥of␈α≡[Green69]␈α≡and␈α≥[Sacerdoti73]␈α≡are␈α≡examples␈α≥of
␈↓ α←␈↓implementationally␈α
implicit␈α
strategies.␈α
 In␈α
both,␈α
the␈α
strategies␈α
are␈α∞realized␈α
as
␈↓ α←␈↓distinct␈α∞sections␈α∞of␈α∞code,␈α∞but␈α∞the␈α
code␈α∞is␈α∞embedded␈α∞in␈α∞the␈α∞control␈α
structure.
␈↓ α←␈↓Green␈α⊂notes␈α⊃that␈α⊂while␈α⊃there␈α⊂is␈α⊂a␈α⊃certain␈α⊂range␈α⊃of␈α⊂control␈α⊃available␈α⊂over
␈↓ α←␈↓strategy␈αif␈αthe␈αuser␈αunderstands␈αprogram␈αparameters␈α(e.g.,␈αthe␈αlevel␈αbound␈αon
␈↓ α←␈↓resolution),␈αmore␈αextensive␈αchanges␈αrequire␈αknowledge␈αof␈α␈↓¬LISP␈↓␈αand␈αan␈αability
␈↓ α←␈↓to recode parts of the system.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α⊗meta-rules␈α⊗are␈α⊗an␈α∃example␈α⊗of␈α⊗explicit␈α⊗representation␈α∃of
␈↓ α←␈↓strategies,␈α
since␈α
they␈α
are␈α
constructs␈α
independent␈α
of␈α
the␈α
control␈α
structure␈α
and
␈↓ α←␈↓represent␈α⊃strategies␈α⊃in␈α⊃much␈α⊃the␈α⊂same␈α⊃way␈α⊃that␈α⊃object-level␈α⊃knowledge␈α⊂is
␈↓ α←␈↓represented by object-level rules.

␈↓"β␈↓ α←␈↓␈↓α7-6-3    Knowledge organization␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αnext␈αdimension␈αconcerns␈αthe␈αorganization␈αof␈αstrategic␈αknowledge
␈↓ α←␈↓and␈α→considers␈α~the␈α→indexing␈α~scheme␈α→around␈α~which␈α→the␈α~strategies␈α→are
␈↓ α←␈↓organized.␈α A␈αnumber␈αof␈α
different␈αapproaches␈αhave␈αbeen␈αused,␈α
among␈αthem
␈↓ α←␈↓are strategies organized around:

␈↓"β␈↓ α←␈↓␈↓ β'␈↓↓KSs␈↓:  In␈α⊗␈↓¬HEARSAY␈α⊗II␈↓,␈α⊗for␈α⊗instance,␈α⊗each␈α⊗KS␈α⊗carries␈α⊗with␈α↔it␈α⊗a
␈↓ α←␈↓␈↓ β'description of the circumstances under which it is most relevant.

␈↓"β␈↓ α←␈↓␈↓ β'␈↓↓Goals␈↓:  As illustrated, meta-rules are associated with specific goals.

␈↓"β␈↓ α←␈↓␈↓ β'␈↓↓Goals␈αin␈α
context␈↓:  Since␈α␈↓¬PLANNER␈↓'s␈α
``advice␈αlist''␈α
of␈αlikely␈α
theorems␈αto
␈↓ α←␈↓␈↓ β'try␈α∞(supplied␈α
by␈α∞the␈α
␈↓	THUSE␈↓␈α∞and␈α
theorem␈α∞base␈α
filter␈α∞constructs)␈α
is
␈↓ α←␈↓␈↓ β'part␈α
of␈α
a␈α
specific␈α
goal␈α
statement,␈α
the␈α
advice␈α
can␈α
be␈α
organized␈α
by
␈↓ α←␈↓␈↓ β'both␈αgoal␈αand␈αcontext.␈α That␈αis,␈αtwo␈α␈↓	THGOAL␈↓␈αstatements␈αmay␈αhave
␈↓ α←␈↓␈↓ β'the␈α
same␈α
goal␈α
pattern␈α
but␈αdifferent␈α
advice␈α
lists␈α
because␈α
they␈αare␈α
in
␈↓ α←␈↓␈↓ β'two different theorems.

␈↓"β␈↓ α←␈↓␈↓ β'␈↓↓Events␈↓:  The␈α⊃␈↓	WHEN␈↓␈α∩statement␈α⊃in␈α⊃␈↓¬QA4␈↓,␈α∩for␈α⊃example,␈α⊃offers␈α∩a␈α⊃rich
␈↓ α←␈↓␈↓ β'syntax␈α↔for␈α↔describing␈α⊗events,␈α↔along␈α↔with␈α↔associated␈α⊗specified
␈↓ α←␈↓␈↓ β'responses.␈α In␈α
general,␈α``event''␈α
can␈αbe␈α
interpreted␈αbroadly␈α
to␈αrefer
␈↓ α←␈↓␈↓ β'to␈α⊂any␈α⊃configuration␈α⊂of␈α⊃the␈α⊂data␈α⊂base␈α⊃allowing␈α⊂strategies␈α⊃to␈α⊂be
␈↓ α←␈↓␈↓ β'organized␈α≠around␈α≠``situations''␈α≠(data␈α≠base␈α≠configurations)␈α~of
␈↓ α←␈↓␈↓ β'varying levels of specificity.

␈↓"β␈↓ α←␈↓␈↓ β?For␈α⊃the␈α⊃purposes␈α⊃of␈α⊃this␈α⊃discussion,␈α⊃it␈α⊃will␈α⊃be␈α⊃useful␈α⊃to␈α⊃make␈α⊂the
␈↓ α←␈↓simple␈α⊂distinction␈α⊂between␈α⊂strategies␈α⊂that␈α⊂are␈α⊂organized␈α⊂around␈α∂individual
␈↓ α←␈↓KSs␈α∂(the␈α∂first␈α∞example;␈α∂they␈α∂will␈α∞be␈α∂referred␈α∂to␈α∞as␈α∂␈↓↓KS-centered␈↓)␈α∂and␈α∞those
␈↓ α←␈↓that␈α
are␈αorganized␈α
around␈αanything␈α
else␈α
(all␈αthe␈α
rest;␈α␈↓↓non-KS-centered␈↓).␈α
 Note
␈↓ α←␈↓␈↓244    STRATEGIES␈↓ 
#7-6␈↓

␈↓"β␈↓ α←␈↓that␈αthose␈αin␈αthe␈αfirst␈αclass␈αare␈α␈↓↓indexed␈↓␈αby␈αKS␈αand␈α␈↓↓refer␈↓␈αto␈αgoals␈αor␈αevents␈αin
␈↓ α←␈↓which␈α∂the␈α∞KS␈α∂is␈α∂useful;␈α∞the␈α∂latter␈α∞are␈α∂␈↓↓indexed␈↓␈α∂around␈α∞goals␈α∂of␈α∂events␈α∞and
␈↓ α←␈↓␈↓↓refer␈↓ to KSs.

␈↓"β␈↓ α←␈↓␈↓α7-6-4    Character␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∪final␈α∪dimension␈α∀offers␈α∪four␈α∪classifications␈α∪for␈α∀describing␈α∪the
␈↓ α←␈↓``character''␈α∀of␈α∀strategy␈α∪content.␈α∀ These␈α∀four␈α∪arise␈α∀from␈α∀two␈α∪independent
␈↓ α←␈↓distinctions␈α⊗(Table␈α⊗)␈α⊗that␈α↔provide␈α⊗a␈α⊗framework␈α⊗for␈α↔understanding␈α⊗the
␈↓ α←␈↓different␈αtypes␈αof␈α
meta-rules␈αexplored␈αearlier.␈α
 They␈αare␈αdescribed␈αinitially␈α
in
␈↓ α←␈↓the␈α
paradigm␈α
of␈α
tree␈α
search␈α
and␈α
then␈α
generalized␈α
to␈αdemonstrate␈α
applicability
␈↓ α←␈↓for␈αa␈αrange␈αof␈αproblem-solving␈αparadigms.␈α Specific␈αexamples␈αillustrate␈αeach
␈↓ α←␈↓classification.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α
modified␈α
breadth-first␈α
tree␈α
search␈α
might␈α
involve␈α
collecting␈α∞all␈α
the
␈↓ α←␈↓nodes␈α⊃that␈α⊂are␈α⊃descendents␈α⊃of␈α⊂the␈α⊃current␈α⊃node␈α⊂and␈α⊃attempting␈α⊃to␈α⊂decide
␈↓ α←␈↓which␈αis␈αthe␈αbest␈αto␈αpursue.␈α Information␈αconcerning␈αthe␈αutility␈αof␈αeach␈αnode
␈↓ α←␈↓might␈α
be␈αof␈α
two␈αforms.␈α
By␈α␈↓↓individual␈α
utility␈↓␈αwe␈α
mean␈αthe␈α
sort␈αtypically␈α
found
␈↓ α←␈↓in␈α∀heuristic␈α∀search␈α∀evaluation␈α∀functions.␈α∀ For␈α∀our␈α∀purposes,␈α∀its␈α∪essential
␈↓ α←␈↓characteristic␈α∩is␈α∩that␈α⊃it␈α∩gives␈α∩a␈α⊃utility␈α∩figure␈α∩for␈α⊃a␈α∩node␈α∩based␈α∩solely␈α⊃on
␈↓ α←␈↓knowledge␈α
about␈α
that␈α
single␈α
node␈α∞and␈α
about␈α
the␈α
current␈α
state␈α
of␈α∞the␈α
world.
␈↓ α←␈↓It does not refer to any of the other nodes.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α∞␈↓↓comparative␈α∂utility␈↓,␈α∞on␈α∂the␈α∞other␈α∞hand,␈α∂specifically␈α∞refers␈α∂to␈α∞more
␈↓ α←␈↓than a single node so that it can offer statements about their relative utility.␈↓
25␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Now␈α
take␈α
a␈α
depth-first␈α
view␈α
of␈α
the␈α
tree␈α
and␈α
imagine␈α
that␈α
the␈α
nodes
␈↓ α←␈↓represent␈α⊗potential␈α⊗invocations␈α⊗of␈α∃backward-chained␈α⊗rules.␈α⊗ As␈α⊗we␈α∃saw
␈↓ α←␈↓earlier,␈α
a␈α
sequence␈α
of␈α
several␈αof␈α
them␈α
a␈α
few␈α
levels␈αdeep␈α
can␈α
be␈α
viewed␈α
as␈αa
␈↓ α←␈↓``line␈α
of␈α
reasoning.'' ␈αWe␈α
can␈α
thus␈α
have␈αstrategies␈α
which␈α
indicate␈α
either␈α``this
␈↓ α←␈↓line␈α∞of␈α∂reasoning␈α∞is␈α∂useful''␈α∞(individual␈α∞utility)␈α∂or␈α∞``this␈α∂line␈α∞of␈α∂reasoning␈α∞is
␈↓ α←␈↓more useful than that one'' (comparative).
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂primary␈α∞advantage␈α∂of␈α∞the␈α∂line-of-reasoning␈α∞type␈α∂strategy␈α∂is␈α∞its
␈↓ α←␈↓ability␈α⊂to␈α⊃direct␈α⊂the␈α⊃system␈α⊂through␈α⊂local␈α⊃minima␈α⊂of␈α⊃evaluation␈α⊂functions.
␈↓ α←␈↓Since␈αone␈αintuitive␈αaspect␈αof␈αintelligence␈αis␈αthe␈αability␈αto␈αperform␈αactions␈αthat
␈↓ α←␈↓may␈α⊂superficially␈α⊂appear␈α⊂counterproductive␈α∂(but␈α⊂which␈α⊂are␈α⊂in␈α⊂fact␈α∂useful),
␈↓ α←␈↓this seems to be an important source of knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?By␈αmaking␈αa␈αfew␈αsimple␈αreplacements,␈αthis␈αscheme␈αcan␈αbe␈αgeneralized
␈↓ α←␈↓to␈αfit␈αmany␈αforms␈αof␈αproblem␈αsolving: ␈αFor␈α``node'',␈αread␈αKS;␈αfor␈α``descendants
␈↓ α←␈↓of␈α
the␈α
current␈α
node,''␈α
read␈α
plausible␈α
knowledge␈α
source␈α
set;␈α
and,␈αfinally,␈α
replace
␈↓ α←␈↓``depth''␈α∞with␈α∞time.␈α∞ Thus,␈α∞instead␈α∞of␈α
speaking␈α∞of␈α∞evaluating␈α∞the␈α∞nodes␈α∞at␈α
a
␈↓ α←␈↓given␈αlevel␈αof␈αthe␈αtree,␈αwe␈αspeak␈αof␈αchoosing␈αa␈αKS␈αfrom␈αthe␈αPKS␈αset;␈α
instead
␈↓ α←␈↓of␈α⊃thinking␈α∩of␈α⊃search␈α∩more␈α⊃than␈α∩one␈α⊃level␈α∩deep,␈α⊃we␈α∩think␈α⊃ahead␈α∩to␈α⊃the
␈↓ α←␈↓invocation of more than one KS.
␈↓"β␈↓ α←␈↓␈↓ β?With␈α∀these␈α∀substitutions,␈α∀we␈α∀can␈α∀also␈α∀take␈α∀another␈α∀look␈α∀at␈α∪each

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[25]␈α∀Note␈α∀that␈α∀while␈α∀it␈α∀is␈α∀possible␈α∀to␈α∀arrive␈α∀at␈α∀such␈α∀a␈α∃comparison␈α∀by
␈↓ α←␈↓examining␈α
individual␈α
utilities,␈α
this␈α
refers␈α
to␈α
strategies␈α
that␈α
compare␈α
two␈α
nodes
␈↓ α←␈↓explicitly.
␈↓ α←␈↓␈↓7-6␈↓ π↑A TAXONOMY, OF SORTS    245␈↓

␈↓"β␈↓ α←␈↓strategy␈α
type␈α
and␈αconsider␈α
examples␈α
from␈αcurrent␈α
systems.␈α
 There␈α
are␈αmeta-
␈↓ α←␈↓rules␈α∂illustrating␈α∞two␈α∂of␈α∂these␈α∞categories:␈α∂␈↓	METARULE001␈↓␈α∞is␈α∂of␈α∂the␈α∞individual
␈↓ α←␈↓utility type, and ␈↓	METARULE002␈↓ indicates comparative utility.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α_are␈α_numerous␈α_examples␈α_in␈α_other␈α_systems␈α_as␈α→well.␈α_ The
␈↓ α←␈↓preconditions␈α⊃of␈α∩the␈α⊃operators␈α⊃in␈α∩␈↓¬HEARSAY␈α⊃II␈↓␈α⊃can␈α∩be␈α⊃viewed␈α∩as␈α⊃individual
␈↓ α←␈↓utility␈α∀statements,␈α∀since␈α∀they␈α∀are␈α∪associated␈α∀with␈α∀an␈α∀individual␈α∀KS␈α∪and
␈↓ α←␈↓describe␈α∪the␈α∪circumstances␈α∩under␈α∪which␈α∪that␈α∩KS␈α∪is␈α∪appropriate,␈α∩without
␈↓ α←␈↓referring to other KSs.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α[Waldinger74]␈αthere␈αare␈αcomparative␈αutility␈αstrategies␈αembodied␈αin
␈↓ α←␈↓the␈α∞␈↓	GOALCLASS␈↓␈α∞statements␈α
which␈α∞specify␈α∞an␈α∞order␈α
for␈α∞the␈α∞use␈α∞of␈α
applicable
␈↓ α←␈↓theorems.␈α
 The␈α∞␈↓¬PLANNER␈↓␈α
example␈α∞above␈α
is␈α
another␈α∞instance␈α
of␈α∞a␈α
comparative
␈↓ α←␈↓utility␈α
strategy;␈α
it␈α
makes␈α
explicit␈α
reference␈α
to␈α
more␈α
than␈α
one␈α
KS␈α
in␈α
order␈α
to
␈↓ α←␈↓specify a relative ordering.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α
early␈α
example␈α
of␈α
a␈α
line-of-reasoning␈α
type␈α
strategy␈α
is␈αmentioned
␈↓ α←␈↓in␈αSamuel's␈αdiscussion␈αof␈αhis␈αchecker␈αprogram␈α[Samuel67],␈αwhere␈αhe␈αterms␈αit
␈↓ α←␈↓a␈α∩``principle␈α∩line.'' ␈α∩The␈α∩macro␈α∩operators␈α∩in␈α∩␈↓¬STRIPS␈↓␈α∩described␈α∪in␈α∩[Fikes72]
␈↓ α←␈↓provide␈α⊂another␈α⊂example,␈α⊂as␈α⊂do␈α⊃meta-rules␈α⊂(although␈α⊂we␈α⊂have␈α⊂not␈α⊃as␈α⊂yet
␈↓ α←␈↓uncovered any examples specific to the medical domain).




␈↓"␈↓ α←␈↓∧                    Breadth                   Depth
␈↓"␈↓ α←␈↓∧            ⊂ααααααααααααααααααααααααααααααααααααααααααααααααα⊃
␈↓"␈↓ α←␈↓∧            ~                                                 ~
␈↓"␈↓ α←␈↓∧Individual  ~   Individual utility        Line of reasoning   ~
␈↓"␈↓ α←␈↓∧            ~                                                 ~
␈↓"␈↓ α←␈↓∧            ~                                                 ~
␈↓"␈↓ α←␈↓∧Comparative ~   Comparative utility       Comparison of lines ~
␈↓"␈↓ α←␈↓∧            ~                             of reasoning        ~
␈↓"␈↓ α←␈↓∧            ~                                                 ~
␈↓"␈↓ α←␈↓∧            %ααααααααααααααααααααααααααααααααααααααααααααααααα$


␈↓"␈↓ α←␈↓α␈↓ ∧YFig. 7-6.    Strategy characteristics.    
␈↓ α←␈↓␈↓246    STRATEGIES␈↓ 
#7-6␈↓



␈↓ α←␈↓α␈↓ ∧DTable 7-3.    The Strategy Taxonomy.    


␈↓"␈↓ α←␈↓∧ααααααααααααπααααααααααααααααααααααααααααααπααααααααααααααααααα
␈↓"␈↓ α←␈↓∧Dimension   ~  Sample points               ~   Example
␈↓"␈↓ α←␈↓∧ααααααααααααβααααααααααααααααααααααααααααααβααααααααααααααααααα
␈↓"␈↓ α←␈↓∧            ~                              ~
␈↓"␈↓ α←␈↓∧Generality  ~                              ~
␈↓"␈↓ α←␈↓∧            ~  Representation independent  ~   Polya
␈↓"␈↓ α←␈↓∧            ~  Domain independent          ~   Green, ABSTRIPS
␈↓"␈↓ α←␈↓∧            ~  Domain specific             ~   Gelernter
␈↓"␈↓ α←␈↓∧            ~  Goal specific               ~   PLANNER
␈↓"␈↓ α←␈↓∧            ~                              ~
␈↓"␈↓ α←␈↓∧            ~                              ~
␈↓"␈↓ α←␈↓∧Explicitness~                              ~
␈↓"␈↓ α←␈↓∧            ~  Conceptually implicit       ~   Howe
␈↓"␈↓ α←␈↓∧            ~  Implementationally implicit ~   Green, ABSTRIPS
␈↓"␈↓ α←␈↓∧            ~  Explicit                    ~   TEIRESIAS
␈↓"␈↓ α←␈↓∧            ~                              ~
␈↓"␈↓ α←␈↓∧            ~                              ~
␈↓"␈↓ α←␈↓∧Organization~                              ~
␈↓"␈↓ α←␈↓∧            ~  KS-centered                 ~   HEARSAY II
␈↓"␈↓ α←␈↓∧            ~  Non-KS-centered,            ~   PLANNER: THUSE
␈↓"␈↓ α←␈↓∧            ~     referral by name         ~
␈↓"␈↓ α←␈↓∧            ~  Non-KS-centered,            ~   PLANNER: THTBF,
␈↓"␈↓ α←␈↓∧            ~     referral by description  ~    TEIRESIAS
␈↓"␈↓ α←␈↓∧            ~                              ~
␈↓"␈↓ α←␈↓∧            ~                              ~
␈↓"␈↓ α←␈↓∧Character   ~                              ~
␈↓"␈↓ α←␈↓∧            ~  Individual, breadth         ~   HEARSAY II
␈↓"␈↓ α←␈↓∧            ~  Comparative, breadth        ~   PLANNER, QA4
␈↓"␈↓ α←␈↓∧            ~  Individual, depth           ~   STRIPS macro ops
␈↓"␈↓ α←␈↓∧            ~  Comparative, depth          ~   - - - - -
␈↓"␈↓ α←␈↓∧            ~                              ~
␈↓"␈↓ α←␈↓∧αααααααααααα∀αααααααααααααααααααααααααααααα∀ααααααααααααααααααα
␈↓ α←␈↓␈↓7-7␈↓ ¬rLIMITATIONS OF THE GENERAL FORMALISM    247␈↓

␈↓"β␈↓ α←␈↓␈↓α7-7    LIMITATIONS OF THE GENERAL FORMALISM␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∃are␈α∃several␈α∃fundamental␈α∃problems␈α∃associated␈α∃with␈α∃a␈α∀fully
␈↓ α←␈↓general␈α∞implementation␈α∞of␈α∞the␈α∂meta-rule␈α∞framework: ␈α∞First,␈α∞it␈α∞is␈α∂not␈α∞always
␈↓ α←␈↓possible␈αto␈α
organize␈αa␈αprogram␈α
as␈αa␈αcollection␈α
of␈αdiscrete␈α
knowledge␈αsources;
␈↓ α←␈↓second,␈αthere␈αis␈αa␈αconsiderable␈αintellectual␈αeffort␈αrequired␈αto␈αassemble␈αall␈αthe
␈↓ α←␈↓elements␈αnecessary␈αfor␈αthe␈αgeneral␈αimplementation;␈αthird,␈αthe␈αscheme␈αwill␈αnot
␈↓ α←␈↓work␈α∞well␈α∞if␈α∞rules␈α∞are␈α∂created␈α∞dynamically␈α∞(i.e.,␈α∞during␈α∞a␈α∂performance␈α∞run);
␈↓ α←␈↓and␈α
finally,␈α
not␈α
all␈α
of␈α
the␈α
overhead␈α
required␈α
for␈α
execution␈α
time␈α
flexibility␈α
can
␈↓ α←␈↓be␈α∞restricted␈α∞to␈α∂the␈α∞background␈α∞and␈α∞may␈α∂impose␈α∞a␈α∞significant␈α∂reduction␈α∞in
␈↓ α←␈↓speed.

␈↓"β␈↓ α←␈↓␈↓α7-7-1    Program organization␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
need␈α∞to␈α
organize␈α∞the␈α
program␈α
as␈α∞a␈α
collection␈α∞of␈α
discrete␈α∞KSs␈α
is
␈↓ α←␈↓perhaps␈α∪the␈α∪most␈α∩fundamental␈α∪limitation␈α∪encountered.␈α∩ While␈α∪it␈α∪may␈α∩be
␈↓ α←␈↓possible␈αto␈αconstrue␈αthe␈αdefinition␈αbroadly␈αenough␈αto␈αfit␈αalmost␈αany␈αprogram,
␈↓ α←␈↓it␈α
is␈αnot␈α
always␈α
an␈αinformative␈α
view.␈α There␈α
seem␈α
to␈αbe␈α
two␈αimportant␈α
factors
␈↓ α←␈↓in␈α∞deciding␈α∞whether,␈α∞and␈α∞how,␈α∂to␈α∞decompose␈α∞a␈α∞program␈α∞into␈α∂distinct␈α∞KSs: 
␈↓ α←␈↓(a)  The␈α∞proposed␈α∞KSs␈α∂should␈α∞be␈α∞self-contained,␈α∞in␈α∂that␈α∞each␈α∞can␈α∂by␈α∞itself
␈↓ α←␈↓make␈αsome␈α
contribution␈αto␈αthe␈α
problem␈αsolution,␈α
and␈α(b)  the␈αproblem␈α
should
␈↓ α←␈↓be␈α
sufficiently␈α
ill-specified␈α
(in␈α
the␈αsense␈α
discussed␈α
in␈α
Section␈α
7-3-1)␈αthat␈α
there
␈↓ α←␈↓is␈αno␈αpredetermined␈αsequence␈αof␈αKS␈αinvocations␈αthat␈αwill␈αsolve␈αthe␈αproblem;
␈↓ α←␈↓instead␈α∪there␈α∪is,␈α∪at␈α∪each␈α∪point,␈α∪some␈α∪indeterminacy␈α∪about␈α∪which␈α∀KS␈α∪to
␈↓ α←␈↓invoke.
␈↓"β␈↓ α←␈↓␈↓ β?If␈α
there␈α
is␈α
but␈α∞a␈α
single␈α
KS␈α
in␈α
the␈α∞program␈α
or␈α
(what␈α
amounts␈α∞to␈α
the
␈↓ α←␈↓same␈α∩thing)␈α∩if␈α∩the␈α∩various␈α∩sources␈α∩of␈α∩knowledge␈α∩are␈α∩so␈α∩intertwined␈α⊃that
␈↓ α←␈↓separating␈α∀them␈α∀is␈α∀impossible,␈α∀then␈α∀we␈α∀clearly␈α∀can't␈α∀meet␈α∃condition␈α∀(a).
␈↓ α←␈↓Similarly,␈α∞if␈α∂there␈α∞is␈α∞an␈α∂algorithmic␈α∞solution␈α∞to␈α∂the␈α∞problem,␈α∞then␈α∂we␈α∞have
␈↓ α←␈↓none␈αof␈αthe␈αindeterminacy␈αsuggested␈αin␈α(b)␈αand␈αdon't␈αneed␈αall␈αthe␈αmachinery
␈↓ α←␈↓outlined above.

␈↓"β␈↓ α←␈↓␈↓α7-7-2    Intellectual difficulty␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αintellectual␈αdifficulty␈α
of␈αimplementing␈αthis␈α
approach␈αarises␈αout␈α
of
␈↓ α←␈↓several␈α⊂factors.␈α⊃ It␈α⊂is␈α⊂not␈α⊃often␈α⊂easy,␈α⊂for␈α⊃example,␈α⊂to␈α⊂describe␈α⊃rather␈α⊂than
␈↓ α←␈↓name␈αa␈αKS.␈α Consider␈αthe␈αsegment␈αof␈αcode␈αin␈αSection␈α7-5␈αthat␈αmight␈αappear
␈↓ α←␈↓in␈αa␈α␈↓¬PLANNER␈↓␈αversion␈αof␈αa␈αpoker␈αgame.␈αThe␈αstrategy␈αthere␈αis␈αto␈αtry␈αbluffing,␈αif
␈↓ α←␈↓that␈α∞doesn't␈α∞work␈α∂draw␈α∞three␈α∞cards,␈α∂and␈α∞if␈α∞all␈α∂else␈α∞fails,␈α∞cheat.␈α∂ To␈α∞replace
␈↓ α←␈↓that␈α with␈α a␈α description-oriented␈α approach␈α requires␈α a␈α!more␈α basic
␈↓ α←␈↓understanding␈α⊗of␈α⊗the␈α⊗domain␈α⊗and␈α⊗of␈α⊗the␈α⊗strategy␈α⊗being␈α↔expressed.␈α⊗A
␈↓ α←␈↓rewritten␈αversion␈α
might␈αsay␈α
something␈αlike␈α``first␈α
use␈αany␈α
psychological␈αploys
␈↓ α←␈↓to␈α⊃discourage␈α⊃the␈α⊃competition,␈α⊃then␈α⊂try␈α⊃something␈α⊃that␈α⊃will␈α⊃improve␈α⊂your
␈↓ α←␈↓hand,␈α
and␈α
finally␈αdo␈α
anything␈α
that␈αwill␈α
make␈α
sure␈α
you␈αwin.'' ␈α
Each␈α
of␈αthese␈α
is
␈↓ α←␈↓a␈α
more␈α
general␈αdescription␈α
of␈α
the␈αentire␈α
class␈α
of␈αtheorems␈α
from␈α
which␈αeach␈α
of
␈↓ α←␈↓the␈α
ones␈αnamed␈α
might␈αhave␈α
been␈αdrawn.␈α
 This␈αis␈α
not␈αoften␈α
easy␈α
to␈αproduce
␈↓ α←␈↓and␈α
makes␈α
certain␈α
assumptions␈α
about␈α
the␈α
nature␈α
of␈α
the␈α
application␈αdomain.
␈↓ α←␈↓In␈α∂particular,␈α∂the␈α∂domain␈α⊂must␈α∂be␈α∂sufficiently␈α∂formalized␈α⊂(or␈α∂formalizable)
␈↓ α←␈↓␈↓248    STRATEGIES␈↓ 
#7-7␈↓

␈↓"β␈↓ α←␈↓that␈α⊃it␈α⊃is␈α⊃a␈α⊃reasonable␈α⊃task␈α⊂to␈α⊃look␈α⊃for␈α⊃meta-level␈α⊃primitives.␈α⊃ There␈α⊂are
␈↓ α←␈↓clearly domains for which this is not true.
␈↓"β␈↓ α←␈↓␈↓ β?Despite␈α⊂these␈α⊂difficulties,␈α⊃the␈α⊂direct␈α⊂benefits␈α⊂are␈α⊃considerable.␈α⊂ The
␈↓ α←␈↓resulting␈α∂flexibility␈α∞can␈α∂be␈α∞an␈α∂important␈α∂tool␈α∞in␈α∂the␈α∞construction␈α∂of␈α∂a␈α∞very
␈↓ α←␈↓large␈α
system.␈α There␈α
are,␈αas␈α
well,␈αindirect␈α
benefits.␈α We␈α
have␈αnoted␈α
that␈αthis
␈↓ α←␈↓approach␈α∞requires␈α∞a␈α∞more␈α∞basic␈α
understanding␈α∞of␈α∞the␈α∞domain␈α∞and␈α
strategy.
␈↓ α←␈↓This␈αin␈α
itself␈αcan␈α
be␈αbeneficial␈αand␈α
lead␈αto␈α
further␈αinsights.␈α More␈α
important,
␈↓ α←␈↓however,␈α∀the␈α∪resulting␈α∀strategy␈α∪offers␈α∀a␈α∪more␈α∀complete␈α∪statement␈α∀of␈α∪the
␈↓ α←␈↓relevant␈α
knowledge.␈α It␈α
makes␈αmore␈α
evident␈αwhat␈α
it␈αis␈α
the␈α
strategy␈αexpresses
␈↓ α←␈↓and␈α
why␈αthe␈α
system␈α
performs␈αas␈α
it␈α
does.␈α It␈α
is␈α
thus␈αanother␈α
example␈α
of␈αone␈α
of
␈↓ α←␈↓our␈α≤recurring␈α≠themes--the␈α≤benefits␈α≤of␈α≠the␈α≤explicit␈α≤representation␈α≠of
␈↓ α←␈↓knowledge.

␈↓"β␈↓ α←␈↓␈↓α7-7-3    Dynamic rule creation␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α→compile-time␈α→vs.␈α~execution-time␈α→distinction␈α→and␈α~the␈α→``pre-
␈↓ α←␈↓compilation''␈α∪of␈α∀descriptions␈α∪into␈α∪corresponding␈α∀sets␈α∪of␈α∪rule␈α∀names␈α∪offer
␈↓ α←␈↓efficiency␈α
only␈αif␈α
the␈αaddition␈α
of␈α
rules␈αto␈α
the␈αsystem␈α
is␈αrestricted␈α
to␈α
the␈αtime
␈↓ α←␈↓between␈αperformance␈αruns.␈α The␈αapproach␈αwe␈αhave␈αbeen␈αdescribing␈αimposes
␈↓ α←␈↓a␈α
significant␈αcost␈α
on␈αthe␈α
attempt␈αto␈α
add␈α
rules␈αto␈α
the␈αknowledge␈α
base␈αwhile␈α
the
␈↓ α←␈↓program␈α∪is␈α∩in␈α∪the␈α∪midst␈α∩of␈α∪a␈α∩performance␈α∪run.␈α∪ At␈α∩worst,␈α∪it␈α∪means␈α∩all
␈↓ α←␈↓descriptions␈α
in␈αstrategies␈α
have␈αto␈α
be␈αrun␈α
``interpreted,''␈αconstantly␈α
recomputing
␈↓ α←␈↓the␈α∞applicable␈α
set␈α∞of␈α∞rules.␈α
 A␈α∞slightly␈α∞more␈α
efficient␈α∞solution␈α∞might␈α
employ
␈↓ α←␈↓an␈α∂incremental␈α⊂compilation,␈α∂which␈α∂recomputed␈α⊂all␈α∂the␈α⊂strategy␈α∂descriptions
␈↓ α←␈↓with␈α⊃reference␈α⊃to␈α⊃the␈α⊃single␈α⊃new␈α⊃rule.␈α⊃ Which␈α⊃of␈α⊃these␈α⊃is␈α⊃more␈α⊃desirable
␈↓ α←␈↓depends␈α
on␈αthe␈α
individual␈αsystem,␈α
but␈α
in␈αeither␈α
case␈αthe␈α
computational␈αcost␈α
is
␈↓ α←␈↓significant.

␈↓"β␈↓ α←␈↓␈↓α7-7-4    Overhead␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∀final␈α∀problem␈α∀is␈α∀the␈α∀cost␈α∀in␈α∀speed␈α∀of␈α∀great␈α∃execution␈α∀time
␈↓ α←␈↓flexibility.␈α∀ If,␈α∀at␈α∀every␈α∀point␈α∀where␈α∀a␈α∀KS␈α∀must␈α∀be␈α∀chosen,␈α∃the␈α∀system
␈↓ α←␈↓performs␈αa␈αfull-blown␈αre-analysis--one␈αthat␈αincludes␈αdeciding␈αonce␈αagain␈αon
␈↓ α←␈↓a␈α
methodology,␈α
deciding␈α
how␈α
to␈α∞decide,␈α
and␈α
so␈α
on--nothing␈α
would␈α∞ever␈α
get
␈↓ α←␈↓accomplished.␈α
 Such␈α
continual␈α
complete␈α∞reappraisal␈α
is␈α
in␈α
most␈α
cases␈α∞a␈α
waste
␈↓ α←␈↓of␈αtime␈αand␈αshould␈αnot␈α
be␈αdone.␈α The␈αquestion␈αis,␈α
thus,␈αhow␈αto␈αinvoke␈αall␈α
(or
␈↓ α←␈↓any)␈αof␈α
this␈α``deciding-how-to-decide''␈α
procedure␈αonly␈α
when␈αnecessary.␈α
 There
␈↓ α←␈↓are several possibilities.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
first,␈αand␈α
perhaps␈αthe␈α
most␈αeffective,␈α
technique␈αbecomes␈α
clear␈αif
␈↓ α←␈↓we␈αrecall␈α
our␈αdefinition␈α
of␈αill␈αstructured␈α
problems.␈α If␈α
a␈αproblem␈αis␈α
completely
␈↓ α←␈↓ill␈α∞structured,␈α∞then␈α
the␈α∞use␈α∞of␈α
the␈α∞fully␈α∞general␈α
machinery␈α∞is␈α∞warranted.␈α
To
␈↓ α←␈↓the␈α⊂extent␈α⊂that␈α⊂it␈α⊂is␈α⊂well␈α⊂structured,␈α⊂the␈α⊂machinery␈α⊂is␈α⊂inappropriate.␈α∂ The
␈↓ α←␈↓point␈αthen␈αis␈αto␈αcombine␈αboth␈αof␈αthese␈αand␈αto␈αallow␈αeach␈αto␈αhelp␈αanswer␈αthe
␈↓ α←␈↓fundamental␈α∞question␈α∞of␈α∞which␈α∞KS␈α∞to␈α∞invoke␈α∞next.␈α∞ The␈α∞answer␈α∂may␈α∞then
␈↓ α←␈↓come␈α∞from␈α∂either␈α∞the␈α∞fully␈α∂general␈α∞``deciding-how-to-decide''␈α∂mechanism␈α∞or
␈↓ α←␈↓from␈α⊃one␈α⊃of␈α⊂the␈α⊃structures␈α⊃that␈α⊂expresses␈α⊃the␈α⊃known␈α⊃connections␈α⊂between
␈↓ α←␈↓KSs.
␈↓ α←␈↓␈↓7-7␈↓ ¬rLIMITATIONS OF THE GENERAL FORMALISM    249␈↓

␈↓"β␈↓ α←␈↓␈↓ β?The␈α∩result␈α∩may␈α∪be␈α∩pictured␈α∩as␈α∩a␈α∪construction␈α∩in␈α∩which␈α∪the␈α∩fully
␈↓ α←␈↓general␈αre-analysis␈α
mechanisms␈αplay␈αthe␈α
role␈αof␈αmortar,␈α
occasionally␈αholding
␈↓ α←␈↓together␈α
bricks␈α
(structured␈α
subproblems),␈α
at␈αother␈α
times␈α
filling␈α
in␈α
molds␈α(the
␈↓ α←␈↓super-structure).␈α⊗ The␈α⊗size␈α⊗of␈α⊗the␈α⊗molds␈α⊗and␈α⊗bricks␈α⊗may␈α↔vary␈α⊗widely,
␈↓ α←␈↓corresponding␈α∩to␈α∩the␈α∩extent␈α∩of␈α∩understanding␈α∩of␈α∩different␈α∩aspects␈α∩of␈α∩the
␈↓ α←␈↓problem.␈αThe␈αissue␈αthus␈αbecomes␈αwhere␈αand␈αhow␈αthe␈αproblem␈αis␈αcut␈αup.␈α As
␈↓ α←␈↓much␈α
of␈α∞it␈α
as␈α∞possible␈α
should␈α
be␈α∞embodied␈α
in␈α∞structures␈α
that␈α∞reflect␈α
known
␈↓ α←␈↓interrelationships,␈α↔leaving␈α↔the␈α↔most␈α↔general␈α↔technique␈α↔as␈α↔a␈α⊗mechanism
␈↓ α←␈↓available␈α
either␈α
to␈αdirect␈α
the␈α
use␈αof␈α
those␈α
structures␈αor␈α
to␈α
fill␈αin␈α
if␈α
none␈αare
␈↓ α←␈↓available.
␈↓"β␈↓ α←␈↓␈↓ β?To␈αmake␈αthis␈αabstract␈αpicture␈αmore␈αconcrete,␈αconsider␈αthe␈αplace␈αof␈α
the
␈↓ α←␈↓meta-rules␈α∞in␈α∞␈↓¬MYCIN␈↓'s␈α∞problem-solving␈α∞formalism.␈α∞ The␈α∂major␈α∞superstructure
␈↓ α←␈↓there␈αis␈αthe␈αbackward␈αchaining␈αof␈αindividual␈αproduction␈αrules␈α
producing␈αan
␈↓ α←␈↓and/or␈αgoal␈α
tree.␈α Within␈αthis␈α
framework,␈αmeta-rules␈αare␈α
used␈αto␈α
help␈αguide
␈↓ α←␈↓the␈α∞search␈α∞through␈α∞the␈α∞tree.␈α
 This␈α∞arrangement␈α∞constrains␈α∞the␈α∞general␈α
(and
␈↓ α←␈↓expensive)␈α
decision␈α∞mechanism␈α
to␈α
a␈α∞collection␈α
of␈α
well-specified␈α∞locations.␈α
 It
␈↓ α←␈↓can thus prove useful even when applied in only a part of the overall system.
␈↓"β␈↓ α←␈↓␈↓ β?At␈α∞the␈α∞beginning␈α∞of␈α∞this␈α∞chapter␈α∞we␈α∞claimed␈α∞that␈α∞the␈α∞framework␈α
of
␈↓ α←␈↓knowledge␈αorganization␈αsuggested␈αby␈αa␈αfairly␈αgeneral␈αconception␈αof␈αstrategies
␈↓ α←␈↓offered␈α⊂a␈α⊂useful␈α⊂and␈α⊃illuminating␈α⊂device␈α⊂for␈α⊂knowledge␈α⊃explication.␈α⊂ This
␈↓ α←␈↓point␈α
can␈α
be␈α
illustrated␈αby␈α
this␈α
same␈α
issue␈αof␈α
problem␈α
structure.␈α
 It␈αwas␈α
noted,
␈↓ α←␈↓above,␈α⊂that␈α⊂the␈α⊂unrestricted␈α⊂use␈α⊂of␈α∂the␈α⊂most␈α⊂general␈α⊂formalism␈α⊂is␈α⊂in␈α∂most
␈↓ α←␈↓cases␈α
a␈α
waste␈α
of␈α
time.␈α
 If␈α
the␈αreader␈α
agrees,␈α
then␈α
the␈α
question␈α
is␈α
␈↓↓how␈α
do␈αyou
␈↓ α←␈↓↓know?␈↓ ␈α∂Or␈α∂more␈α∂precisely,␈α∂␈↓↓what␈↓␈α∂is␈α∂it␈α∂that␈α∂you␈α∂know␈α∂that␈α∂suggests␈α⊂that␈α∂the
␈↓ α←␈↓complete␈α
re-analysis␈α
is␈α
unwarranted? ␈α
What␈α
kind␈α
of␈α
(or␈α
how␈α
many)␈αfailures
␈↓ α←␈↓would␈α
it␈α
take␈α
before␈αyou␈α
began␈α
to␈α
change␈α
your␈αmind? ␈α
This␈α
is␈α
just␈α
the␈αsort␈α
of
␈↓ α←␈↓knowledge␈α∩that␈α⊃can␈α∩be␈α∩formalized␈α⊃to␈α∩provide␈α⊃direction␈α∩for␈α∩the␈α⊃problem-
␈↓ α←␈↓solving␈α⊂process.␈α∂ It␈α⊂may␈α⊂possibly␈α∂be␈α⊂embedded␈α⊂in␈α∂one␈α⊂of␈α⊂the␈α∂conventional
␈↓ α←␈↓approaches,␈α⊃or␈α⊃perhaps␈α⊃in␈α⊂a␈α⊃line-of-reasoning␈α⊃style␈α⊃second-order␈α⊂strategy:
␈↓ α←␈↓``Don't␈α∀reconsider␈α∀the␈α∪methodology␈α∀until␈α∀the␈α∪next␈α∀three␈α∀KSs␈α∀have␈α∪been
␈↓ α←␈↓invoked using the current scheme.''
␈↓"β␈↓ α←␈↓␈↓ β?There␈αis␈α
one␈αadditional␈α
approach␈αto␈αall␈α
of␈αthis␈α
that␈αmay␈αprove␈α
useful.
␈↓ α←␈↓In␈α
an␈αanalogy␈α
to␈αhuman␈α
behavior,␈αwe␈α
might␈αorganize␈α
the␈αsystem␈α
to␈α
allow␈αit
␈↓ α←␈↓to␈α
fail␈α
``upwards''␈α
through␈α
different␈αlevels␈α
of␈α
generality␈α
in␈α
its␈α
attempts␈αto␈α
solve
␈↓ α←␈↓the␈α∞problem.␈↓
26␈↓␈α∂ ␈α∞Consider␈α∂the␈α∞problem␈α∞of␈α∂turning␈α∞on␈α∂a␈α∞lamp␈α∂and␈α∞noticing
␈↓ α←␈↓that␈α
the␈α∞light␈α
does␈α∞not␈α
come␈α∞on.␈α
We␈α
might␈α∞try␈α
a␈α∞sequence␈α
of␈α∞solutions␈α
that
␈↓ α←␈↓looked like:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?see if the lamp is plugged in,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?try replacing the bulb,
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?see if the lamp switch is o.k.,
␈↓"β␈↓ α←␈↓␈↓ ββ(d)␈↓ β?see if there is a fuse blown, or
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[26]␈α≠A␈α≠similar␈α≠concept␈α≠of␈α~successive␈α≠levels␈α≠of␈α≠failure␈α≠is␈α≠found␈α~in
␈↓ α←␈↓[Winograd75].␈α⊃The␈α⊃basic␈α⊃idea␈α⊃is␈α⊃also␈α⊂inherent␈α⊃in␈α⊃the␈α⊃view␈α⊃of␈α⊃AI␈α⊃as␈α⊂the
␈↓ α←␈↓``science of weak methods'' [Newell69].
␈↓ α←␈↓␈↓250    STRATEGIES␈↓ 
#7-7␈↓

␈↓"β␈↓ α←␈↓␈↓ ββ(e)␈↓ β?find out if there has been a power failure.


␈↓ α←␈↓Each␈α⊃failure␈α⊃causes␈α⊃successively␈α⊃fewer␈α⊃assumptions␈α⊃about␈α⊃the␈α⊃state␈α⊃of␈α⊂the
␈↓ α←␈↓world␈α∂to␈α∞be␈α∂taken␈α∂for␈α∞granted␈α∂and␈α∞causes␈α∂a␈α∂more␈α∞general␈α∂approach␈α∂to␈α∞the
␈↓ α←␈↓solution.␈α∪ Analogously,␈α∪a␈α∩program␈α∪failing␈α∪in␈α∩its␈α∪heuristic␈α∪search␈α∪via␈α∩hill
␈↓ α←␈↓climbing␈α∩might␈α∩first␈α∩re-evaluate␈α∩its␈α∪distance␈α∩metric,␈α∩then␈α∩the␈α∩use␈α∪of␈α∩hill
␈↓ α←␈↓climbing,␈αand␈αfinally␈αthe␈αuse␈αof␈αheuristic␈αsearch.␈α In␈αthis␈αcase␈αthe␈αgeneral␈αre-
␈↓ α←␈↓evaluation mechanism is invoked only where necessary.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α
can␈α∞comment,␈α
too,␈α∞on␈α
the␈α∞investment␈α
needed␈α∞for␈α
implementation
␈↓ α←␈↓of␈α⊂this␈α⊂methodology.␈α∂Since␈α⊂the␈α⊂benefits␈α⊂(and␈α∂costs)␈α⊂are␈α⊂all␈α⊂incremental,␈α∂we
␈↓ α←␈↓need␈α∀not␈α∀implement␈α∀the␈α∀entire␈α∪formalism␈α∀to␈α∀get␈α∀any␈α∀payoff.␈α∀ Even␈α∪the
␈↓ α←␈↓smallest steps toward it are useful.
␈↓"β␈↓ α←␈↓␈↓ β?But␈α
is␈α
all␈α
this␈αreally␈α
necessary?␈α
 That␈α
is,␈αis␈α
all␈α
of␈α
this␈αalways␈α
relevant? 
␈↓ α←␈↓Clearly␈α∂not.␈α∂There␈α∂surely␈α∂are␈α∂times␈α∂when␈α∂systems␈α∂are␈α∂written␈α∂to␈α⊂achieve␈α∂a
␈↓ α←␈↓very␈α
well-specified␈α∞purpose␈α
and␈α∞will␈α
not␈α∞necessarily␈α
profit␈α∞from␈α
considering
␈↓ α←␈↓the␈α∀task␈α∀in␈α∀its␈α∀full␈α∀generality,␈α∀as␈α∀suggested␈α∀here.␈α∀ It␈α∀must␈α∃therefore␈α∀be
␈↓ α←␈↓desirable to have such generality as a part of the system being constructed.
␈↓"β␈↓ α←␈↓␈↓ β?There␈αare,␈αhowever,␈αtwo␈αpoints␈αto␈αbe␈αconsidered.␈α First,␈αeven␈αa␈αsystem
␈↓ α←␈↓that␈α⊂is␈α⊃not␈α⊂designed␈α⊃for␈α⊂extensive␈α⊂generality␈α⊃may␈α⊂be␈α⊃structured␈α⊂somewhat
␈↓ α←␈↓more␈α∞cleanly␈α∞if␈α∞the␈α∞knowledge␈α∞is␈α
viewed␈α∞along␈α∞the␈α∞lines␈α∞we␈α∞have␈α
suggested
␈↓ α←␈↓(e.g.,␈α
in␈α
Green's␈α
system).␈α Second,␈α
it␈α
is␈α
rare␈αthat␈α
all␈α
the␈α
necessary␈αcapabilities
␈↓ α←␈↓of␈α∪a␈α∪system␈α∪can␈α∪be␈α∪foreseen␈α∪in␈α∪advance,␈α∪and␈α∪redesign␈α∪is␈α∪a␈α∩well-known
␈↓ α←␈↓occupational␈αhazard.␈α Note␈αthat␈αthe␈αdesign␈αand␈αconstruction␈αof␈α
any␈αprogram
␈↓ α←␈↓(even␈α
for␈α
a␈α
well-understood␈α
task␈α
domain)␈α
is␈α
itself␈α
an␈α
ill␈α
structured␈α
problem.
␈↓ α←␈↓Thus,␈α
it␈αmay␈α
still␈αprove␈α
useful␈αto␈α
attempt␈αto␈α
be␈αgeneral,␈α
even␈αin␈α
the␈αcontext
␈↓ α←␈↓of a well-defined task area.

␈↓"β␈↓ α←␈↓␈↓α7-8    SUMMARY␈↓
␈↓"β␈↓ α←␈↓␈↓ β?This␈α∂chapter␈α⊂has␈α∂proposed␈α∂a␈α⊂general␈α∂framework␈α⊂for␈α∂understanding
␈↓ α←␈↓strategies␈α⊃as␈α⊂a␈α⊃form␈α⊂of␈α⊃meta-level␈α⊂knowledge␈α⊃and␈α⊂has␈α⊃suggested␈α⊃that␈α⊂any
␈↓ α←␈↓decision␈α∩concerning␈α∪the␈α∩use␈α∪of␈α∩knowledge␈α∩in␈α∪problem␈α∩solving␈α∪should␈α∩be
␈↓ α←␈↓viewed␈αas␈αa␈αstrategic␈αchoice.␈α It␈αexplored␈αhow␈αthis␈αframework␈αcan␈αhelp␈αmake
␈↓ α←␈↓clear␈αthe␈αbody␈αof␈αknowledge␈αcontained␈αin␈αa␈αprogram␈αand␈αshowed␈αhow␈αit␈αcan
␈↓ α←␈↓contribute to both representation design and organization.
␈↓"β␈↓ α←␈↓␈↓ β?Meta-rules␈α⊂were␈α⊂described␈α⊂in␈α⊂detail␈α⊂as␈α⊂one␈α⊂instance␈α⊂of␈α⊃the␈α⊂general
␈↓ α←␈↓ideas␈α⊂outlined␈α⊃in␈α⊂the␈α⊃framework.␈α⊂ We␈α⊂have␈α⊃seen␈α⊂how␈α⊃they␈α⊂help␈α⊃to␈α⊂guide
␈↓ α←␈↓problem-solving␈α⊃performance␈α⊃and␈α∩have␈α⊃explored␈α⊃the␈α⊃range␈α∩of␈α⊃knowledge
␈↓ α←␈↓that␈αthey␈αcan␈αbe␈αused␈αto␈αrepresent.␈α Explanation␈αof␈αmeta-rules␈αwas␈αshown␈αto
␈↓ α←␈↓be␈αpossible␈αwith␈α
a␈αstraightforward␈αextension␈α
of␈αthe␈αfacilities␈α
for␈αobject-level
␈↓ α←␈↓rules; steps toward interactive acquisition of meta-rules were also considered.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α∪saw␈α∀that␈α∪meta-rules␈α∀refer␈α∪to␈α∀object-level␈α∪rules␈α∀by␈α∪describing
␈↓ α←␈↓(rather␈α
than␈αnaming)␈α
them␈αand␈α
that␈αthey␈α
effect␈αthis␈α
reference␈α
by␈αexamining
␈↓ α←␈↓their␈α
content␈α
directly.␈α This␈α
led␈α
to␈α
a␈αconsideration␈α
of␈α
the␈αdifferent␈α
approaches
␈↓ α←␈↓to␈α∂invoking␈α∂a␈α∂knowledge␈α∂source␈α∂and␈α∂offered␈α∂a␈α∂perspective␈α∂on␈α∂the␈α∞relevant
␈↓ α←␈↓␈↓7-8␈↓ 	βSUMMARY    251␈↓

␈↓"β␈↓ α←␈↓factors␈α∞involved.␈α∞ Content-directed␈α
invocation␈α∞was␈α∞shown␈α
to␈α∞offer␈α∞a␈α
handle
␈↓ α←␈↓on␈αKSs␈α
that␈αprovides␈αvalidity,␈α
expressiveness,␈αand␈α
an␈αexplicit␈αspecification␈α
of
␈↓ α←␈↓invocation␈α∪criteria.␈α∀ While␈α∪it␈α∀requires␈α∪a␈α∀significant␈α∪initial␈α∀investment␈α∪of
␈↓ α←␈↓effort,␈α∩it␈α∩can␈α∩provide␈α∩a␈α∩useful␈α∩level␈α∩of␈α∩flexibility␈α∩for␈α∩systems␈α∪with␈α∩large
␈↓ α←␈↓knowledge bases subject to frequent changes.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α∞a␈α∞taxonomy␈α∞of␈α∞strategy␈α∞types␈α∞was␈α∞reviewed,␈α∞offering␈α∞a␈α
basis
␈↓ α←␈↓for␈α~comparing␈α≠meta-rules␈α~to␈α≠other␈α~forms␈α≠of␈α~strategy␈α≠encoding.␈α~ An
␈↓ α←␈↓examination␈α∂of␈α∂the␈α∂impact␈α∂of␈α∂the␈α∂different␈α∂forms␈α∂of␈α∂encoding␈α∂led␈α∂to␈α∞some
␈↓ α←␈↓observations␈α⊂about␈α⊃organizing␈α⊂knowledge␈α⊂in␈α⊃ways␈α⊂that␈α⊂help␈α⊃make␈α⊂systems
␈↓ α←␈↓both easier to construct and more flexible in the face of changes.
␈↓ α←␈↓␈↓␈↓ 
α    253␈↓




␈↓"β␈↓ α←␈↓␈↓ ε∃␈↓αChapter 8



␈↓"β␈↓ α←␈↓α␈↓ ∧u␈↓λCONCLUSIONS␈↓α


␈↓"β␈↓ α←␈↓α␈↓ ¬
summary; the themes revisited









␈↓"β␈↓ α←␈↓␈↓ ¬GAlas,␈α⊂how␈α⊂terrible␈α⊂is␈α⊂wisdom␈α⊂when␈α⊂it␈α⊃brings␈α⊂no
␈↓ α←␈↓␈↓ ¬Gprofit␈α	to␈α	the␈α	man␈α	that's␈αλwise!␈α	 This␈α	I␈α	knew␈α	well,␈αλbut
␈↓ α←␈↓␈↓ ¬Ghad forgotten it; else I would not have come here.
␈↓"β␈↓ α←␈↓␈↓ 	∨lines 312-313

␈↓"β␈↓ α←␈↓␈↓α8-1    INTRODUCTION␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αlast␈αfive␈αchapters␈αhave␈αexplored␈αa␈αnumber␈αof␈αdifferent␈αproblems
␈↓ α←␈↓encountered␈α⊃in␈α⊃building␈α⊂␈↓¬TEIRESIAS␈↓,␈α⊃a␈α⊃program␈α⊂intended␈α⊃to␈α⊃establish␈α⊃a␈α⊂link
␈↓ α←␈↓between␈α∞a␈α∞human␈α∂expert␈α∞and␈α∞a␈α∂computer␈α∞consultant.␈α∞ Those␈α∂problems␈α∞and
␈↓ α←␈↓solutions␈αare␈αreviewed␈αbriefly␈αhere␈αand␈αconsidered␈αin␈αthe␈αlight␈αof␈αmeta-level
␈↓ α←␈↓knowledge␈α
as␈α
a␈α
tool␈α∞for␈α
knowledge␈α
base␈α
construction,␈α
maintenance,␈α∞and␈α
use.
␈↓ α←␈↓This␈α∪is␈α∪followed␈α∪by␈α∪a␈α∪discussion␈α∪of␈α∪some␈α∪of␈α∪the␈α∪global␈α∪limitations␈α∩and
␈↓ α←␈↓shortcomings␈α⊂that␈α∂arise␈α⊂from␈α∂the␈α⊂attempt␈α∂to␈α⊂link␈α∂the␈α⊂expert␈α⊂and␈α∂program.
␈↓ α←␈↓Finally,␈α∀we␈α∀return␈α∀to␈α∀the␈α∀alternative␈α∀set␈α∀of␈α∀themes␈α∀listed␈α∀in␈α∃chapter␈α∀1,
␈↓ α←␈↓reconsidering␈αsome␈αof␈αthis␈α
work␈αin␈αthose␈αterms␈αand␈α
using␈αthem␈αas␈αa␈αbasis␈α
for
␈↓ α←␈↓speculation about future directions.

␈↓"β␈↓ α←␈↓␈↓α8-2    REVIEW OF MAJOR ISSUES␈↓

␈↓"β␈↓ α←␈↓␈↓α8-2-1    Forms of meta-level knowledge␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Fig.␈α∪8-1␈α∪reviews␈α∪the␈α∪three␈α∪major␈α∪forms␈α∪of␈α∪meta-level␈α∩knowledge
␈↓ α←␈↓developed␈α⊃in␈α⊃previous␈α⊂chapters.␈α⊃ The␈α⊃rule␈α⊂models␈α⊃describe␈α⊃the␈α⊃␈↓↓content␈↓␈α⊂of
␈↓ α←␈↓inference␈α
rules␈α
in␈α
the␈α
knowledge␈αbase,␈α
making␈α
clear␈α
the␈α
global␈α
trends␈αin␈α
those
␈↓ α←␈↓rules␈α∩and␈α∩providing␈α∪useful␈α∩assistance␈α∩in␈α∪acquiring␈α∩new␈α∩rules.␈α∪ The␈α∩data
␈↓ α←␈↓structure␈αschemata␈αdescribe␈α
the␈α␈↓↓structure␈↓␈αof␈α
the␈αconceptual␈αprimitives␈αused␈α
in
␈↓ α←␈↓␈↓254    CONCLUSIONS␈↓ 
#8-2␈↓

␈↓"β␈↓ α←␈↓expressing␈α⊃rules␈α∩and␈α⊃offer␈α∩a␈α⊃basis␈α∩for␈α⊃the␈α∩acquisition␈α⊃of␈α∩new␈α⊃primitives.
␈↓ α←␈↓Finally,␈αthe␈αmeta-rules␈αdescribe␈αhow␈αto␈α␈↓↓use␈↓␈αobject-level␈αrules␈αand␈αfunction␈αas
␈↓ α←␈↓strategies to guide invocation of those rules.
␈↓"β␈↓ α←␈↓␈↓ β?One␈αadditional,␈α
less␈αdeveloped␈αsource␈α
of␈αmeta-level␈αknowledge␈α
is␈αthe
␈↓ α←␈↓function␈α∃templates␈α∃(chapter␈α∃2),␈α∃which␈α∀describe␈α∃part␈α∃of␈α∃the␈α∃structure␈α∀of
␈↓ α←␈↓inference rules and, like the rule models, are used in acquiring new rules.


␈↓"β␈↓ α←␈↓	   FORM              DESCRIBES               USE


␈↓"β␈↓ α←␈↓	Rule models         CONTENT of           Acquisition of
␈↓"β␈↓ α←␈↓	                    inference rules      new inference rules

␈↓"β␈↓ α←␈↓	Data structure      STRUCTURE of         Acquisition of
␈↓"β␈↓ α←␈↓	schemata            conceptual           new conceptual
␈↓"β␈↓ α←␈↓	                    primitives           primitives

␈↓"β␈↓ α←␈↓	Meta-rules          USE of object-       Guide use of
␈↓"β␈↓ α←␈↓	                    level rules          object-level knowledge



␈↓"β␈↓ α←␈↓α␈↓ βcFig. 8-1.    The three forms of meta-level knowledge.    


␈↓"β␈↓ α←␈↓␈↓α8-2-2    Explanation␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Chapter␈α_3␈α_discussed␈α_efforts␈α_to␈α_enable␈α_␈↓¬TEIRESIAS␈↓␈α_to␈α_explain␈α_the
␈↓ α←␈↓reasoning␈αof␈αa␈αperformance␈αprogram␈αto␈αaudiences␈αthat␈αrange␈αfrom␈αan␈αexpert
␈↓ α←␈↓acquainted␈α∂with␈α∂the␈α∂program␈α∂to␈α∞a␈α∂student␈α∂with␈α∂minimal␈α∂experience␈α∂in␈α∞the
␈↓ α←␈↓field.␈α→ The␈α~basic␈α→steps␈α→in␈α~building␈α→the␈α→explanation␈α~facility␈α→involved
␈↓ α←␈↓(a)  determining␈α
the␈α
program␈α
operation␈αthat␈α
was␈α
to␈α
be␈α
considered␈αprimitive,
␈↓ α←␈↓(b)  augmenting␈α↔the␈α↔performance␈α↔program␈α↔to␈α↔leave␈α↔behind␈α↔a␈α_trace␈α↔of
␈↓ α←␈↓behavior␈α
in␈α
terms␈α
of␈α
this␈α
operation,␈α
(c)  finding␈α
a␈α
framework␈α
in␈α
which␈αthat
␈↓ α←␈↓trace␈α
could␈α
be␈α
understood,␈α
and␈α
(d)  designing␈α
a␈α
program␈α
that␈α
would␈αallow␈α
the
␈↓ α←␈↓user␈α
to␈α
examine␈α
the␈α
behavior␈α
trace␈α
in␈α
terms␈α
of␈α
the␈α
framework.␈α
 This␈α
supplied
␈↓ α←␈↓a␈α∂mechanism␈α⊂for␈α∂exploring␈α∂past␈α⊂actions;␈α∂future␈α∂actions␈α⊂of␈α∂the␈α⊂system␈α∂were
␈↓ α←␈↓explored␈α∂by␈α∞means␈α∂of␈α∂code␈α∞in␈α∂the␈α∂explanation␈α∞system␈α∂which␈α∂simulated␈α∞the
␈↓ α←␈↓control structure of the performance program.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α
approach␈α
to␈α
explanation␈α
relies␈α
on␈α
the␈α
assumption␈α
that␈α
a␈α
recap␈α
of
␈↓ α←␈↓program␈α∂operations␈α∂can␈α∂be␈α∂a␈α∂reasonable␈α∂basis␈α∂for␈α∂generating␈α∂explanations.
␈↓ α←␈↓This␈αwas␈αseen␈αto␈αsuggest␈αthe␈αareas␈αof␈αgreatest␈αapplicability␈αas␈αwell␈αas␈αimply␈αa
␈↓ α←␈↓number␈αof␈αlimitations.␈α In␈αparticular,␈αit␈αseems␈αbest␈αsuited␈αto␈αprograms␈αwhose
␈↓ α←␈↓primary␈α⊃mode␈α⊃of␈α⊃computation␈α∩is␈α⊃symbolic␈α⊃reasoning␈α⊃rather␈α∩than␈α⊃numeric
␈↓ α←␈↓operations.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∀primary␈α∪shortcomings␈α∀in␈α∪the␈α∀current␈α∪implementation␈α∀were␈α∪a
␈↓ α←␈↓␈↓8-2␈↓ π<REVIEW OF MAJOR ISSUES    255␈↓

␈↓"β␈↓ α←␈↓failure␈α∞to␈α∞factor␈α∞in␈α
the␈α∞state␈α∞of␈α∞the␈α∞viewer's␈α
knowledge,␈α∞a␈α∞lack␈α∞of␈α∞a␈α
general
␈↓ α←␈↓notion␈α∞of␈α
what␈α∞an␈α
explanation␈α∞is,␈α
and␈α∞a␈α
lack␈α∞of␈α
ability␈α∞to␈α∞represent␈α
control
␈↓ α←␈↓structures.␈α∞ The␈α
latter␈α∞becomes␈α
particularly␈α∞evident␈α
when␈α∞changes␈α∞are␈α
made
␈↓ α←␈↓to␈αthe␈αcontrol␈αstructure␈αof␈αthe␈αperformance␈αprogram,␈αsince␈αthis␈αoften␈αrequires
␈↓ α←␈↓significant␈α⊂changes␈α⊂to␈α⊂the␈α∂explanation␈α⊂program.␈α⊂ We␈α⊂speculated␈α⊂about␈α∂the
␈↓ α←␈↓possibility␈α≠of␈α≠a␈α≤representation␈α≠of␈α≠control␈α≠structures␈α≤that␈α≠emphasized
␈↓ α←␈↓intentional␈αinformation,␈αwhich,␈αcombined␈αwith␈αa␈αformalization␈αof␈αthe␈α
concept
␈↓ α←␈↓of␈α⊃explanation,␈α⊂might␈α⊃make␈α⊃the␈α⊂system␈α⊃far␈α⊃more␈α⊂flexible␈α⊃and␈α⊃general.␈α⊂ It
␈↓ α←␈↓might,␈αideally,␈αbe␈αpossible␈αfor␈αthe␈αsystem␈αto␈αexamine␈αits␈αown␈αcode,␈αgenerating
␈↓ α←␈↓explanations on the basis of what it found there.

␈↓"β␈↓ α←␈↓␈↓α8-2-3    Knowledge acquisition␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Knowledge␈α∃acquisition␈α∃was␈α∃described␈α∃as␈α∃a␈α∃process␈α∃of␈α∀interactive
␈↓ α←␈↓transfer␈α⊂of␈α⊂expertise␈α∂from␈α⊂an␈α⊂expert␈α⊂to␈α∂a␈α⊂performance␈α⊂program,␈α⊂in␈α∂which
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓'s␈αtask␈αwas␈α
to␈α``listen''␈αas␈αattentively␈α
and␈αintelligently␈αas␈αpossible.␈α
 The
␈↓ α←␈↓process␈αwas␈αset␈αin␈αthe␈αcontext␈αof␈αa␈αshortcoming␈αin␈αthe␈αknowledge␈αbase,␈αas␈αan
␈↓ α←␈↓aid␈α
to␈α
both␈αthe␈α
expert␈α
and␈αthe␈α
system.␈α
 This␈αcontext␈α
provides␈α
the␈αexpert␈α
with
␈↓ α←␈↓a␈α∞useful␈α∞organization␈α∞and␈α∞focus.␈α∞ He␈α
is␈α∞not␈α∞simply␈α∞asked␈α∞to␈α∞describe␈α∞all␈α
he
␈↓ α←␈↓knows␈α⊃about␈α⊃a␈α∩domain.␈α⊃ He␈α⊃is␈α⊃instead␈α∩faced␈α⊃with␈α⊃a␈α∩specific␈α⊃consultation
␈↓ α←␈↓whose␈αresults␈α
he␈αfinds␈α
incorrect␈αand␈α
has␈αavailable␈αto␈α
him␈αa␈α
set␈αof␈α
tools␈αthat
␈↓ α←␈↓will␈α∩allow␈α⊃him␈α∩to␈α⊃uncover␈α∩the␈α⊃extent␈α∩of␈α⊃the␈α∩system's␈α⊃knowledge␈α∩and␈α⊃the
␈↓ α←␈↓rationale␈α
behind␈α∞its␈α
performance.␈α∞ His␈α
task␈α∞is␈α
then␈α∞to␈α
specify␈α∞the␈α
particular
␈↓ α←␈↓difference␈αbetween␈αthe␈αsystem's␈αknowledge␈αand␈αhis␈αown␈αthat␈αaccounts␈αfor␈αthe
␈↓ α←␈↓discrepancy␈αin␈αresults.␈α The␈αsystem␈αrelies␈αon␈αthe␈αcontext␈αof␈αthe␈αerror␈αto␈αform
␈↓ α←␈↓a␈α⊃set␈α⊃of␈α⊂expectations␈α⊃about␈α⊃the␈α⊂character␈α⊃of␈α⊃the␈α⊂information␈α⊃that␈α⊃will␈α⊂be
␈↓ α←␈↓forthcoming.␈α This␈αleads␈αto␈αbetter␈αcomprehension␈αof␈αthe␈αexpert's␈αstatement␈αof
␈↓ α←␈↓that␈αinformation␈αand␈αprovides␈αa␈αnumber␈αof␈αchecks␈αon␈αits␈αcontent␈αthat␈αinsure
␈↓ α←␈↓it␈α
will␈α
in␈α
fact␈α
repair␈α
the␈α
problem␈α
discovered.␈α
 In␈α
a␈α
single␈α
phrase,␈α
␈↓↓interactive
␈↓ α←␈↓↓transfer␈α⊂of␈α∂expertise␈α⊂in␈α∂the␈α⊂context␈α∂of␈α⊂a␈α∂shortcoming␈α⊂in␈α∂the␈α⊂knowledge␈α∂base␈↓
␈↓ α←␈↓characterizes␈α
our␈α
approach␈α
to␈α
the␈α
problem␈α
and␈α
suggests␈α
the␈α
source␈α
of␈αmany
␈↓ α←␈↓of the system's abilities.

␈↓"β␈↓ α←␈↓␈↓αAcquiring new rules␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?New␈α∨rule␈α∨acquisition␈α∨was␈α∨seen␈α∨in␈α∨terms␈α of␈α∨model-directed
␈↓ α←␈↓understanding␈αand␈αa␈αrecognition-oriented␈αapproach␈αto␈αcomprehension.␈α This
␈↓ α←␈↓means␈αthat␈α
the␈αsystem␈α
has␈αsome␈α
model␈αof␈α
the␈αcontent␈α
of␈αthe␈α
signal␈αit␈αis␈α
trying
␈↓ α←␈↓to␈α∞interpret␈α∞and␈α
uses␈α∞this␈α∞to␈α
help␈α∞constrain␈α∞the␈α
set␈α∞of␈α∞interpretations␈α∞it␈α
will
␈↓ α←␈↓consider.␈α⊂ In␈α⊂our␈α∂case,␈α⊂the␈α⊂model␈α∂took␈α⊂the␈α⊂form␈α∂of␈α⊂rule␈α⊂models,␈α∂constructs
␈↓ α←␈↓which␈α
offer␈α
a␈α
picture␈α
of␈α
a␈α
``typical''␈αrule␈α
of␈α
a␈α
given␈α
type.␈α
 These␈α
models␈αare
␈↓ α←␈↓assembled␈αby␈αthe␈α
system␈αitself,␈αusing␈αa␈α
primitive,␈αstatistically␈αoriented␈αform␈α
of
␈↓ α←␈↓concept␈αformation␈αto␈αproduce␈αabstract␈αdescriptions␈αof␈αsyntactic␈αregularities␈αin
␈↓ α←␈↓subsets of rules.
␈↓"β␈↓ α←␈↓␈↓ β?As␈α
noted,␈α∞the␈α
context␈α∞provided␈α
by␈α∞the␈α
process␈α∞of␈α
tracking␈α∞down␈α
the
␈↓ α←␈↓error␈α↔in␈α_the␈α↔knowledge␈α↔base␈α_makes␈α↔it␈α↔possible␈α_for␈α↔␈↓¬TEIRESIAS␈↓␈α_to␈α↔form
␈↓ α←␈↓expectations␈α⊂about␈α⊂the␈α⊂character␈α⊃of␈α⊂the␈α⊂new␈α⊂rule.␈α⊂ These␈α⊃expectations␈α⊂are
␈↓ α←␈↓␈↓256    CONCLUSIONS␈↓ 
#8-2␈↓

␈↓"β␈↓ α←␈↓expressed␈αby␈α
selecting␈αa␈α
specific␈αrule␈αmodel.␈α
 The␈αtext␈α
of␈αthe␈α
new␈αrule␈αis␈α
then
␈↓ α←␈↓allowed␈α∂to␈α∂suggest␈α∂interpretations␈α∂(a␈α∂bottom-up,␈α∂data-directed␈α⊂process),␈α∂but
␈↓ α←␈↓these␈α∂are␈α⊂constrained␈α∂and␈α∂evaluated␈α⊂for␈α∂likely␈α∂validity␈α⊂by␈α∂reference␈α⊂to␈α∂the
␈↓ α←␈↓rule␈αmodel␈α
(a␈αtop-down,␈α
hypothesis-driven␈αprocess).␈α It␈α
is␈αthe␈α
intersection␈αof
␈↓ α←␈↓these␈α
two␈α
information␈α
sources,␈αapproaching␈α
the␈α
task␈α
from␈αdifferent␈α
directions,
␈↓ α←␈↓that is responsible for much of the system's performance.
␈↓"β␈↓ α←␈↓␈↓ β?Further␈α↔application␈α↔of␈α↔the␈α⊗model-directed␈α↔formalism␈α↔is␈α↔seen␈α⊗in
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓'s␈α∞ability␈α∞to␈α∞``second␈α
guess''␈α∞the␈α∞expert.␈α∞ Since␈α
it␈α∞has␈α∞a␈α∞model␈α∞of␈α
the
␈↓ α←␈↓knowledge␈α⊂base--the␈α∂rule␈α⊂models--it␈α∂can␈α⊂tell␈α∂when␈α⊂something␈α∂``fits''␈α⊂in␈α∂the
␈↓ α←␈↓knowledge␈αbase.␈α
 It␈αis␈αthe␈α
occurrence␈αof␈α
a␈αpartial␈αmatch␈α
between␈αthe␈αnew␈α
rule
␈↓ α←␈↓and␈αthe␈αselected␈αrule␈αmodel␈αthat␈αprompts␈α␈↓¬TEIRESIAS␈↓␈αto␈αmake␈αsuggestions␈αto␈αthe
␈↓ α←␈↓expert.␈α⊃ This␈α⊃idea␈α⊃of␈α⊂an␈α⊃unmet␈α⊃expectation␈α⊃is␈α⊂not␈α⊃specific␈α⊃to␈α⊃the␈α⊂current
␈↓ α←␈↓organization␈αor␈αstructure␈αof␈αthe␈αrule␈αmodels␈αand␈αcan␈αbe␈αgeneralized␈αto␈αcover
␈↓ α←␈↓any aspects of a representation about which expectations can be formed.
␈↓"β␈↓ α←␈↓␈↓ β?Several␈α∞implications␈α∞were␈α∞seen␈α∞to␈α∞follow␈α∞from␈α∞the␈α∞fact␈α∞that␈α
␈↓¬TEIRESIAS␈↓
␈↓ α←␈↓constructs␈α∪the␈α∩rule␈α∪models␈α∩from␈α∪current␈α∩contents␈α∪of␈α∩the␈α∪knowledge␈α∩base.
␈↓ α←␈↓Since␈α∂the␈α⊂process␈α∂is␈α⊂automated,␈α∂the␈α⊂expert␈α∂never␈α⊂has␈α∂to␈α⊂enter␈α∂a␈α⊂model␈α∂by
␈↓ α←␈↓hand;␈α∂he␈α∂may␈α∂even␈α∂be␈α∞unaware␈α∂of␈α∂their␈α∂existence.␈α∂ Moreover,␈α∂unlike␈α∞most
␈↓ α←␈↓other␈α
model-based␈α
systems,␈α
new␈α
models␈α
are␈α
constructed␈α
on␈α
the␈α
basis␈α
of␈αpast
␈↓ α←␈↓experience,␈α∞since␈α∞rules␈α∂learned␈α∞previously␈α∞are␈α∂used␈α∞in␈α∞forming␈α∂new␈α∞models.
␈↓ α←␈↓Since␈α
the␈α
models␈α
are␈α
updated␈α
as␈α
each␈α
new␈α
rule␈α
enters␈α
the␈α
knowledge␈α
base,␈α
the
␈↓ α←␈↓model␈α⊂set␈α⊂is␈α∂kept␈α⊂current,␈α⊂evolving␈α⊂with␈α∂the␈α⊂growing␈α⊂knowledge␈α⊂base␈α∂and
␈↓ α←␈↓reflecting the shifting patterns it contains.
␈↓"β␈↓ α←␈↓␈↓ β?Other␈α⊂implications␈α∂follow␈α⊂from␈α∂the␈α⊂fact␈α∂that␈α⊂these␈α∂models␈α⊂give␈α∂the
␈↓ α←␈↓system␈α∂an␈α∂abstract␈α∂picture␈α∞of␈α∂its␈α∂own␈α∂knowledge␈α∞base.␈α∂ It␈α∂means␈α∂that,␈α∂in␈α∞a
␈↓ α←␈↓rudimentary␈α∞way,␈α∞the␈α∞system␈α∞``knows␈α∂what␈α∞it␈α∞knows,␈α∞and␈α∞knows␈α∞where␈α∂it␈α∞is
␈↓ α←␈↓ignorant.'' ␈α
It␈α∞can␈α
answer␈α∞questions␈α
about␈α∞the␈α
content␈α∞of␈α
its␈α∞knowledge␈α
base
␈↓ α←␈↓by␈α∃``reading''␈α∃a␈α∃rule␈α∃model,␈α∃giving␈α∃a␈α∃picture␈α∃of␈α∃global␈α∃structure␈α∃of␈α∃its
␈↓ α←␈↓knowledge␈α∞about␈α∞a␈α∞topic.␈α
 Since␈α∞the␈α∞models␈α∞are␈α
ordered␈α∞on␈α∞the␈α∞basis␈α∞of␈α
an
␈↓ α←␈↓empirically␈α∩defined␈α∪``strength,''␈α∩the␈α∪system␈α∩can␈α∪also␈α∩give␈α∪some␈α∩indications
␈↓ α←␈↓about possible gaps in its knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α∩the␈α∪coupling␈α∩of␈α∩model␈α∪formation␈α∩with␈α∪the␈α∩model-directed
␈↓ α←␈↓understanding␈αprocess␈α
offers␈αa␈αnovel␈α
form␈αof␈αclosed-loop␈α
behavior.␈αExisting
␈↓ α←␈↓rule␈αmodels␈αare␈α
used␈αto␈αguide␈α
the␈αacquisition␈αprocess,␈α
the␈αnew␈αrule␈α
is␈αadded
␈↓ α←␈↓to␈α⊗the␈α⊗knowledge␈α⊗base,␈α⊗and␈α⊗the␈α⊗relevant␈α⊗rule␈α⊗models␈α↔are␈α⊗recomputed.
␈↓ α←␈↓Performance␈α∞of␈α∞the␈α∞acquisition␈α
routines␈α∞may␈α∞thus␈α∞conceivably␈α∞be␈α
improved
␈↓ α←␈↓on the very next rule.
␈↓"β␈↓ α←␈↓␈↓ β?In␈α∩summary,␈α∩␈↓¬TEIRESIAS␈↓␈α⊃constructs␈α∩models␈α∩of␈α⊃the␈α∩knowledge␈α∩base;␈α⊃it
␈↓ α←␈↓updates␈α∪those␈α∪models␈α∪in␈α∪response␈α∩to␈α∪changes,␈α∪keeping␈α∪them␈α∪an␈α∩accurate
␈↓ α←␈↓reflection␈αof␈αthe␈αcurrent␈αknowledge␈αbase;␈αand␈αit␈αthen␈αuses␈αthem␈αto␈αaid␈αin␈αthe
␈↓ α←␈↓acquisition of new knowledge.

␈↓"β␈↓ α←␈↓␈↓αAcquiring new conceptual primitives␈↓    
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂primary␈α∞issue␈α∂here␈α∞is␈α∂the␈α∞representation␈α∂and␈α∞use␈α∂of␈α∞knowledge
␈↓ α←␈↓about␈α⊗representations.␈α↔ The␈α⊗schemata␈α⊗and␈α↔associated␈α⊗structures␈α↔offer␈α⊗a
␈↓ α←␈↓␈↓8-2␈↓ π<REVIEW OF MAJOR ISSUES    257␈↓

␈↓"β␈↓ α←␈↓language␈α∪for␈α∪the␈α∪expression␈α∪of␈α∪the␈α∪knowledge␈α∪and␈α∪a␈α∪framework␈α∪for␈α∪its
␈↓ α←␈↓organization.␈α
 There␈α
are␈α∞three␈α
levels␈α
to␈α∞that␈α
organization:␈α
(␈↓↓i␈↓)␈α∞the␈α
individual
␈↓ α←␈↓schema␈αis␈α
the␈αfundamental␈α
unit␈αof␈α
organization␈αand␈α
is␈αa␈αrecord-like␈α
structure
␈↓ α←␈↓that␈α∩provides␈α⊃the␈α∩basis␈α∩for␈α⊃assembling␈α∩a␈α∩variety␈α⊃of␈α∩information␈α∩about␈α⊃a
␈↓ α←␈↓particular␈α~representation;␈α~(␈↓↓ii␈↓)␈α~the␈α~schema␈α~network␈α~is␈α~a␈α→generalization
␈↓ α←␈↓hierarchy␈α↔that␈α↔indicates␈α⊗the␈α↔existing␈α↔categories␈α⊗of␈α↔data␈α↔structures␈α⊗and
␈↓ α←␈↓relationships␈αbetween␈αthem;␈αand␈α(␈↓↓iii␈↓)␈αthe␈αslotnames␈αand␈αslotexperts␈αthat␈αmake
␈↓ α←␈↓up␈α≥a␈α≡schema␈α≥deal␈α≥with␈α≡specific␈α≥representation␈α≥conventions␈α≡at␈α≥the
␈↓ α←␈↓programming language level.
␈↓"β␈↓ α←␈↓␈↓ β?Unlike␈α⊃standard␈α⊃records,␈α⊃however,␈α⊃the␈α⊃schemata␈α⊃and␈α⊃all␈α⊂associated
␈↓ α←␈↓structures␈α∞are␈α∞a␈α
part␈α∞of␈α∞the␈α
system␈α∞itself␈α∞and␈α
are␈α∞available␈α∞for␈α
examination
␈↓ α←␈↓and␈α∞reference.␈α∂ They␈α∞also␈α∂have␈α∞the␈α∞ability␈α∂to␈α∞describe␈α∂a␈α∞certain␈α∂amount␈α∞of
␈↓ α←␈↓variability in structure description.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α→process␈α_of␈α→acquiring␈α→a␈α_new␈α→conceptual␈α→primitive␈α_strongly
␈↓ α←␈↓resembles␈αthe␈αcreation␈αof␈αa␈αnew␈α
instance␈αof␈αa␈αrecord,␈αbut␈αhas␈α
been␈αextended
␈↓ α←␈↓in␈αseveral␈α
ways.␈α It␈α
has␈αbeen␈α
made␈αinteractive,␈α
to␈αallow␈α
the␈αexpert␈α
to␈αsupply
␈↓ α←␈↓information␈α
about␈αthe␈α
domain;␈αthe␈α
dialog␈αis␈α
couched␈αin␈α
high-level␈α
terms,␈αto
␈↓ α←␈↓make␈α
it␈α
comprehensible␈α
to␈α
a␈α
nonprogrammer;␈α
and␈α
the␈α
whole␈α
process␈αhas␈α
been
␈↓ α←␈↓made␈α
as␈α
easy␈α
and␈α``intelligent''␈α
as␈α
possible,␈α
to␈αease␈α
the␈α
task␈α
of␈αassembling␈α
large
␈↓ α←␈↓amounts of knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α≠approach␈α≠involves␈α≠viewing␈α≠representational␈α≠primitives␈α~as
␈↓ α←␈↓extended␈α⊂data␈α⊂types␈α⊂and␈α⊃constructing␈α⊂the␈α⊂appropriate␈α⊂schema␈α⊂for␈α⊃each␈α⊂of
␈↓ α←␈↓them.␈α∪ That␈α∪is,␈α∪the␈α∪language␈α∪for␈α∪describing␈α∪representations␈α∪was␈α∪used␈α∪to
␈↓ α←␈↓formalize␈α≤a␈α≠range␈α≤of␈α≤information␈α≠about␈α≤the␈α≤performance␈α≠program's
␈↓ α←␈↓representations.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∪generality␈α∪of␈α∪this␈α∪approach␈α∪results␈α∪from␈α∪a␈α∪stratification␈α∩and
␈↓ α←␈↓isolation␈α
of␈α
different␈α
varieties␈α
of␈α
knowledge␈α
at␈α
different␈α
levels: ␈α
Instances␈αof
␈↓ α←␈↓individual␈α⊂schemata␈α∂form␈α⊂the␈α∂collection␈α⊂of␈α∂domain-specific␈α⊂knowledge;␈α∂the
␈↓ α←␈↓schemata␈α∩themselves␈α∩define␈α∪a␈α∩base␈α∩of␈α∪representation-specific␈α∩information;
␈↓ α←␈↓while␈α∪the␈α∪schema-schema␈α∪supplies␈α∪a␈α∪small␈α∪foundation␈α∪of␈α∩representation-
␈↓ α←␈↓independent␈α
knowledge.␈α
 This␈α
stratification␈α
makes␈α
it␈α
possible␈α
for␈α
the␈α
system
␈↓ α←␈↓to␈α∩acquire␈α∩both␈α∩new␈α∩instances␈α∩of␈α∩existing␈α∩representations␈α∩(as␈α∩in␈α⊃learning
␈↓ α←␈↓about␈αa␈αnew␈αorganism)␈α
and␈αnew␈αtypes␈αof␈α
representation␈α(as␈αin␈αthe␈α
acquisition
␈↓ α←␈↓of a new schema), using a single formalism and a single body of code.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α⊂it␈α⊂was␈α⊂noted␈α⊂that␈α⊂the␈α⊂same␈α⊂motivation␈α⊂was␈α⊂responsible␈α⊂for
␈↓ α←␈↓both␈α
the␈α
schemata␈α
and␈α
the␈α
recursive␈α
application␈α
of␈α
the␈α
idea␈α
to␈α∞produce␈α
the
␈↓ α←␈↓schema-schema.␈α∞ The␈α∞schemata␈α∞were␈α∞designed␈α∞to␈α∞automate␈α∞the␈α∂handling␈α∞of
␈↓ α←␈↓the␈α
large␈α
number␈αof␈α
details␈α
involved␈α
in␈αthe␈α
creation␈α
and␈α
management␈αof␈α
data
␈↓ α←␈↓structures.␈α∀ But␈α∀they␈α∀themselves␈α∀were␈α∀sufficiently␈α∀complex,␈α∃detailed␈α∀data
␈↓ α←␈↓structures␈α∞that␈α∞it␈α∞was␈α∞useful␈α∞to␈α∞have␈α∞a␈α∞similar␈α∞device␈α∞for␈α∞their␈α
construction
␈↓ α←␈↓and␈α
management.␈α
 This␈α
resulted␈α
in␈α
the␈α
creation␈α
of␈α
the␈α∞schema-schema,␈α
and
␈↓ α←␈↓it,␈α∃along␈α∃with␈α∃a␈α∃small␈α⊗body␈α∃of␈α∃associated␈α∃structures,␈α∃forms␈α∃a␈α⊗body␈α∃of
␈↓ α←␈↓representation-independent␈αknowledge␈αfrom␈α
which␈αa␈αknowledge␈αbase␈α
can␈αbe
␈↓ α←␈↓constructed.
␈↓ α←␈↓␈↓258    CONCLUSIONS␈↓ 
#8-2␈↓

␈↓"β␈↓ α←␈↓␈↓α8-2-4    Strategies␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αfinal␈αform␈αof␈αmeta-level␈αknowledge␈αexplored␈αwas␈αthe␈α
concept␈αof
␈↓ α←␈↓a␈α
strategy,␈α
defined␈αas␈α
knowledge␈α
about␈α
the␈αuse␈α
of␈α
knowledge.␈α This␈α
definition
␈↓ α←␈↓was␈α∞extended␈α∞to␈α
include␈α∞the␈α∞possibility␈α
of␈α∞an␈α∞arbitrary␈α
number␈α∞of␈α∞levels␈α
of
␈↓ α←␈↓strategies,␈α∞each␈α∂of␈α∞which␈α∞can␈α∂direct␈α∞the␈α∂use␈α∞of␈α∞the␈α∂information␈α∞at␈α∂the␈α∞next
␈↓ α←␈↓lower␈α∀level.␈α∃ We␈α∀considered␈α∃the␈α∀possible␈α∀building␈α∃blocks␈α∀for␈α∃a␈α∀strategy
␈↓ α←␈↓language␈α
and␈αspeculated␈α
about␈α
the␈αsource␈α
of␈αthe␈α
conceptual␈α
primitives␈αfrom
␈↓ α←␈↓which␈αsuch␈αa␈αlanguage␈αmight␈αbe␈αbuilt.␈α The␈αresulting␈αframework␈αwas␈αseen␈αto
␈↓ α←␈↓offer␈α⊂a␈α⊂reasonably␈α⊂general␈α∂view,␈α⊂one␈α⊂that␈α⊂can␈α∂help␈α⊂to␈α⊂organize␈α⊂and␈α∂make
␈↓ α←␈↓explicit␈α∃strategy␈α∃knowledge␈α∃that␈α⊗is␈α∃otherwise␈α∃often␈α∃embedded␈α⊗subtly␈α∃in
␈↓ α←␈↓program code.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α∂explored␈α∂the␈α∂character␈α∂of␈α∂problems␈α∂for␈α∂which␈α∂this␈α⊂approach␈α∂is
␈↓ α←␈↓useful,␈α∩noted␈α∪that␈α∩it␈α∪is␈α∩most␈α∩applicable␈α∪to␈α∩what␈α∪are␈α∩called␈α∪ill␈α∩structured
␈↓ α←␈↓problems␈αand␈α
found␈αthat␈α
it␈αoffers␈α
the␈αgreatest␈α
advantage␈αfor␈α
programs␈αwith
␈↓ α←␈↓large knowledge bases subject to frequent change.
␈↓"β␈↓ α←␈↓␈↓ β?Meta-rules␈αwere␈αdescribed␈αas␈αone␈αexample␈αof␈αthese␈αideas.␈α They␈αwere
␈↓ α←␈↓seen␈α∂to␈α∂offer␈α∂a␈α∂convenient␈α∂mechanism␈α∂for␈α∂the␈α∂expression␈α∂of␈α⊂strategies␈α∂and
␈↓ α←␈↓proved␈α→capable␈α→of␈α~guiding␈α→program␈α→performance␈α~without␈α→introducing
␈↓ α←␈↓unreasonable␈α→overhead.␈α_ The␈α→organization␈α_of␈α→knowledge␈α→they␈α_provide
␈↓ α←␈↓appears␈α⊂novel,␈α∂in␈α⊂guiding␈α⊂heuristic␈α∂search␈α⊂without␈α⊂being␈α∂part␈α⊂of␈α⊂a␈α∂search
␈↓ α←␈↓algorithm.␈α∪ The␈α∪search␈α∪routine␈α∪itself␈α∪is␈α∪very␈α∪simple;␈α∪the␈α∪``intelligence''␈α∪is
␈↓ α←␈↓organized␈αaround␈αand␈αstored␈αin␈αthe␈α
goal␈αtree␈αitself.␈α This␈αmeans␈αthat␈α
at␈αany
␈↓ α←␈↓point␈α∂where␈α⊂the␈α∂system␈α⊂has␈α∂to␈α⊂choose␈α∂a␈α⊂path␈α∂through␈α⊂the␈α∂tree,␈α⊂there␈α∂may
␈↓ α←␈↓possibly be information stored at that branch point available to guide it.
␈↓"β␈↓ α←␈↓␈↓ β?Meta-rules␈αillustrated␈α
several␈αinteresting␈αissues␈α
when␈αconsidered␈αas␈α
an
␈↓ α←␈↓invocation␈α⊂mechanism.␈α∂ First,␈α⊂the␈α∂fact␈α⊂that␈α∂meta-rules␈α⊂refer␈α⊂to␈α∂object-level
␈↓ α←␈↓rules␈αby␈αreference␈αto␈αtheir␈αcontent␈αrather␈αthan␈αby␈αname␈αwas␈αtermed␈αcontent-
␈↓ α←␈↓directed␈α⊗invocation␈α⊗and␈α⊗found␈α⊗to␈α⊗be␈α⊗a␈α⊗generalization␈α⊗of␈α⊗some␈α⊗of␈α∃the
␈↓ α←␈↓traditional␈α∀mechanisms␈α∀of␈α∀knowledge␈α∀source␈α∀invocation.␈α∃ This␈α∀technique
␈↓ α←␈↓offers␈α∃a␈α∀greater␈α∃degree␈α∀of␈α∃expressiveness␈α∀and␈α∃validity␈α∀than␈α∃is␈α∀typically
␈↓ α←␈↓available␈αand␈αwas␈αseen␈αto␈αprovide␈αa␈αhigh␈αlevel␈αof␈αflexibility.␈α
 Second,␈αmeta-
␈↓ α←␈↓rules␈α∂offer␈α⊂a␈α∂framework␈α∂in␈α⊂which␈α∂the␈α∂user␈α⊂can␈α∂define␈α∂his␈α⊂own␈α∂invocation
␈↓ α←␈↓criteria,␈αleading␈αto␈αthe␈αidea␈αof␈αgeneralized␈αinvocation␈αcriteria.␈α This␈αfrees␈αthe
␈↓ α←␈↓programmer␈α
from␈α
the␈αrestriction␈α
of␈α
using␈αonly␈α
those␈α
invocation␈α
criteria␈αthat
␈↓ α←␈↓are␈αpredefined␈αand␈αembedded␈αin␈αthe␈αprogramming␈αlanguage␈αin␈αuse.␈α He␈αcan
␈↓ α←␈↓instead␈α⊂define␈α⊂the␈α⊂set␈α⊂of␈α⊂criteria␈α⊂he␈α⊂wishes␈α⊂to␈α⊂use␈α⊂and␈α⊂can,␈α⊂as␈α⊂well,␈α∂write
␈↓ α←␈↓programs␈αcapable␈αof␈α
choosing␈αfrom␈αamong␈αthat␈α
set␈αat␈αexecution␈αtime.␈α
 Third,
␈↓ α←␈↓the␈α∪presence␈α∪of␈α∪multiple␈α∪levels␈α∪of␈α∪meta-rules␈α∪means␈α∪that␈α∀the␈α∪invocation
␈↓ α←␈↓criteria␈α∞(meta-rules␈α∂at␈α∞one␈α∂level)␈α∞are␈α∞also␈α∂treated␈α∞as␈α∂data␈α∞objects␈α∂(by␈α∞meta-
␈↓ α←␈↓rules␈α
at␈α
the␈α
next␈αhigher␈α
level).␈α
 This␈α
means␈αthat␈α
the␈α
system␈α
has␈α
a␈αprimitive
␈↓ α←␈↓capability to reason about the control structure it should use.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∞limitations␈α∞of␈α∂this␈α∞approach␈α∞and␈α∂its␈α∞costs␈α∞were␈α∂also␈α∞considered.
␈↓ α←␈↓While␈α
it␈α
is␈α
possible␈α
at␈α∞present␈α
to␈α
implement␈α
it␈α
in␈α
simple␈α∞ways,␈α
sophisticated
␈↓ α←␈↓use␈α
requires␈αadvances␈α
in␈αthe␈α
field␈αof␈α
program␈αunderstanding.␈α
 It␈αalso␈α
imposes
␈↓ α←␈↓a␈α∃computational␈α∃overhead.␈α∃ While␈α∃some␈α∃of␈α∃this␈α∃can␈α∃be␈α∃relegated␈α⊗to␈α∃a
␈↓ α←␈↓␈↓8-2␈↓ π<REVIEW OF MAJOR ISSUES    259␈↓

␈↓"β␈↓ α←␈↓background␈α∃pre-compilation␈α∃phase,␈α⊗the␈α∃system␈α∃cannot␈α∃then␈α⊗create␈α∃rules
␈↓ α←␈↓dynamically␈αduring␈αexecution␈αof␈αthe␈αperformance␈αprogram.␈α As␈αnoted␈αabove,
␈↓ α←␈↓the␈α∞approach␈α∞seems␈α∞to␈α∞offer␈α
the␈α∞most␈α∞substantive␈α∞advantages␈α∞when␈α
dealing
␈↓ α←␈↓with␈α∂large␈α∂knowledge␈α∞bases␈α∂with␈α∂numerous␈α∞strategies,␈α∂in␈α∂systems␈α∂subject␈α∞to
␈↓ α←␈↓frequent␈α
change.␈α
 It␈α∞is␈α
thus␈α
appropriate␈α
for␈α∞programs␈α
where␈α
the␈α
size␈α∞of␈α
the
␈↓ α←␈↓knowledge␈αbase␈αand␈αthe␈α
necessity␈αof␈αan␈αincremental␈αapproach␈α
to␈αcompetence
␈↓ α←␈↓require␈αthe␈αability␈αto␈αassemble␈αlarge␈αamounts␈αof␈αknowledge␈αand␈αthe␈αability␈αto
␈↓ α←␈↓make numerous alterations to it.

␈↓"β␈↓ α←␈↓␈↓α8-3    GLOBAL LIMITATIONS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αfundamental␈αproblem␈αwe␈αstarted␈αout␈αto␈αconsider␈αwas␈αthe␈α
creation
␈↓ α←␈↓of␈α∩an␈α∩intelligent␈α∩link␈α∩between␈α⊃a␈α∩domain␈α∩expert␈α∩and␈α∩a␈α⊃high-performance
␈↓ α←␈↓program.␈α∪ This␈α∪provided␈α∩a␈α∪large␈α∪collection␈α∩of␈α∪topics␈α∪for␈α∪discussion,␈α∩and
␈↓ α←␈↓previous␈αsections␈αhave␈αnoted␈αthe␈αshortcomings␈αand␈αlimitations␈αof␈αeach␈αof␈αthe
␈↓ α←␈↓individual␈α
solutions␈α
proposed.␈α
 But␈α
what␈αof␈α
this␈α
basic␈α
notion␈α
of␈α
linking␈αthe
␈↓ α←␈↓expert␈α⊗and␈α⊗program?␈α∃ What␈α⊗limitations␈α⊗might␈α∃there␈α⊗be␈α⊗in␈α⊗that␈α∃whole
␈↓ α←␈↓approach,␈α∀and␈α∀what␈α∀limitations␈α∀may␈α∀arise␈α∀from␈α∀the␈α∃particular␈α∀solutions
␈↓ α←␈↓proposed and implementations described?
␈↓"β␈↓ α←␈↓␈↓ β?First,␈α
the␈α
current␈α∞version␈α
of␈α
␈↓¬TEIRESIAS␈↓␈α
is␈α∞still␈α
quite␈α
rough.␈α
 It␈α∞has␈α
not
␈↓ α←␈↓yet␈α⊂been␈α∂given␈α⊂to␈α∂any␈α⊂users␈α∂for␈α⊂rigorous␈α∂testing.␈α⊂ It␈α∂needs␈α⊂a␈α∂good␈α⊂deal␈α∂of
␈↓ α←␈↓polish␈α∩and␈α⊃more␈α∩attention␈α⊃to␈α∩user␈α⊃convenience␈α∩before␈α⊃it␈α∩becomes␈α∩a␈α⊃truly
␈↓ α←␈↓useful␈α
tool.␈α
 Some␈αsuggestions␈α
along␈α
these␈α
lines␈αwere␈α
noted␈α
in␈αearlier␈α
chapters,
␈↓ α←␈↓but␈α
there␈αis␈α
a␈αmore␈α
basic␈αquestion␈α
concerning␈α
the␈αnature␈α
of␈αthe␈α
tools␈αwe␈α
have
␈↓ α←␈↓assembled.␈α
 Some␈αare␈α
conceptually␈αneat,␈α
but␈αare␈α
they␈αwhat␈α
an␈αexpert␈α
building
␈↓ α←␈↓a large knowledge base would really find most useful?
␈↓"β␈↓ α←␈↓␈↓ β?For␈αinstance,␈αis␈α␈↓¬TEIRESIAS␈↓'s␈αability␈αto␈αsecond␈αguess␈αan␈αexpert's␈αnew␈αrule
␈↓ α←␈↓going␈α∂to␈α⊂help,␈α∂or␈α⊂will␈α∂it␈α⊂only␈α∂get␈α⊂in␈α∂the␈α⊂way?␈α∂ Can␈α⊂it␈α∂be␈α⊂made␈α∂``smarter''?
␈↓ α←␈↓Could␈α
the␈αsystem␈α
learn,␈αfor␈α
example,␈αto␈α
recognize␈αthe␈α
fact␈αthat␈α
the␈α
expert␈αis
␈↓ α←␈↓entering␈α∃a␈α∃whole␈α∃new␈α∃sequence␈α∃of␈α∃rules␈α∃and␈α∃that␈α∃none␈α∃of␈α⊗its␈α∃current
␈↓ α←␈↓expectations␈α∞are␈α∞likely␈α∞to␈α
be␈α∞met?␈↓
1␈↓␈α∞In␈α∞more␈α
general␈α∞terms,␈α∞we␈α∞have␈α
stressed
␈↓ α←␈↓the␈αadvantages␈αof␈αdebugging␈αin␈αcontext,␈αbut␈αare␈αthere␈αways␈αof␈αincreasing␈αthe
␈↓ α←␈↓range␈αof␈αthat␈α
context?␈α Acquisition␈αis␈α
currently␈αa␈αrule-by-rule␈α
process,␈αwhen
␈↓ α←␈↓it really ought to be day-by-day or instructor-by-instructor.
␈↓"β␈↓ α←␈↓␈↓ β?Experience␈α∞with␈α∞a␈α∞number␈α∞of␈α
experts␈α∞may␈α∞provide␈α∞answers␈α∞to␈α
these
␈↓ α←␈↓questions,␈α∞and␈α∞indicate␈α∂where␈α∞tools␈α∞may␈α∞be␈α∂needed␈α∞for␈α∞the␈α∂more␈α∞mundane,
␈↓ α←␈↓but␈α∂more␈α∞immediate,␈α∂problems␈α∞that␈α∂arise␈α∞in␈α∂dealing␈α∞with␈α∂large␈α∂amounts␈α∞of
␈↓ α←␈↓knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
use␈αof␈α
production␈αrules␈α
also␈αposes␈α
certain␈αproblems,␈α
one␈αof␈α
which
␈↓ α←␈↓arises␈α∂from␈α∂the␈α∂impact␈α∂of␈α∂rules␈α∂on␈α∂the␈α∂organization␈α∂of␈α∂knowledge␈α∂and␈α∞the
␈↓ α←␈↓style␈αof␈αprogram␈αwriting,␈αespecially␈αas␈αcompared␈αto␈αa␈αprocedural␈αview.␈α Rules
␈↓ α←␈↓are␈α⊂a␈α⊂reasonably␈α⊂natural␈α⊂and␈α⊂convenient␈α⊂form␈α⊂of␈α⊂knowledge␈α⊂encoding␈α∂for

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[1]␈αIn␈αthis␈αsense␈αthe␈αsystem␈αis␈αcurrently␈αa␈αbit␈αlike␈αan␈αover-eager␈αstudent␈αwho
␈↓ α←␈↓is␈αalways␈α
trying␈αto␈αout-guess␈α
the␈αinstructor.␈α
 There␈αreally␈αought␈α
to␈αbe␈α
a␈αway
␈↓ α←␈↓to tell it to ``keep quiet and listen.''
␈↓ α←␈↓␈↓260    CONCLUSIONS␈↓ 
#8-3␈↓

␈↓"β␈↓ α←␈↓what␈α
may␈α
be␈α∞termed␈α
``single-level''␈α
phenomena--it␈α
is␈α∞easy␈α
to␈α
think␈α∞of␈α
single
␈↓ α←␈↓decisions␈α∃or␈α∃actions␈α∃in␈α⊗terms␈α∃of␈α∃a␈α∃rule.␈α∃ Experience␈α⊗has␈α∃demonstrated,
␈↓ α←␈↓however,␈α⊂that␈α⊃even␈α⊂experts␈α⊂acquainted␈α⊃with␈α⊂the␈α⊂production␈α⊃rule␈α⊂encoding
␈↓ α←␈↓tend␈α∩to␈α⊃think␈α∩of␈α⊃a␈α∩sequence␈α⊃of␈α∩operations␈α⊃in␈α∩procedural␈α⊃terms␈α∩and␈α⊃find
␈↓ α←␈↓flowcharts␈αa␈αconvenient␈αmedium␈αof␈αexpression.␈α While␈αflowcharts␈αcan␈αalways
␈↓ α←␈↓be␈α
converted␈α∞to␈α
an␈α∞equivalent␈α
set␈α
of␈α∞rules,␈α
the␈α∞conversion␈α
is␈α∞nontrivial␈α
and
␈↓ α←␈↓sometimes␈α⊂requires␈α∂reconsidering␈α⊂the␈α⊂knowledge␈α∂being␈α⊂expressed,␈α⊂since␈α∂the
␈↓ α←␈↓two␈α∩methodologies␈α∩offer␈α∩different␈α∩perspectives␈α∩on␈α∩knowledge␈α⊃organization
␈↓ α←␈↓and use.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α∂may␈α⊂as␈α∂well␈α∂bow␈α⊂to␈α∂the␈α∂inevitable,␈α⊂then,␈α∂and␈α∂consider␈α⊂ways␈α∂of
␈↓ α←␈↓formalizing␈α
this␈αtranslation␈α
process.␈α
 How␈αmight␈α
an␈αexpert␈α
describe␈α
a␈αwhole
␈↓ α←␈↓␈↓↓sequence␈↓␈α
of␈α∞decisions,␈α
and␈α∞how␈α
might␈α
the␈α∞system␈α
then␈α∞translate␈α
this␈α∞into␈α
an
␈↓ α←␈↓equivalent␈α∀(unordered)␈α∪␈↓↓set␈↓␈α∀of␈α∀rules?␈↓
2␈↓␈α∪This␈α∀appears␈α∪to␈α∀be␈α∀an␈α∪interesting
␈↓ α←␈↓prospect for further work.
␈↓"β␈↓ α←␈↓␈↓ β?We␈α∂have␈α∂also␈α∂begun␈α∂to␈α∂encounter␈α∂a␈α∂standard␈α∂technical␈α∂problem,␈α∞as
␈↓ α←␈↓the␈α∪combination␈α∪of␈α∀␈↓¬TEIRESIAS␈↓␈α∪and␈α∪␈↓¬MYCIN␈↓␈α∀has␈α∪grown␈α∪beyond␈α∀the␈α∪resources
␈↓ α←␈↓available␈α~in␈α~present␈α≠PDP-10␈α~based␈α~TENEX␈α~␈↓¬INTERLISP␈↓␈α≠systems.␈α~ The
␈↓ α←␈↓combination␈α∂is␈α∞already␈α∂large␈α∞enough␈α∂that␈α∞many␈α∂data␈α∞structures␈α∂have␈α∂to␈α∞be
␈↓ α←␈↓retrieved␈α∂from␈α∂the␈α∂disk␈α∂during␈α∞execution␈α∂rather␈α∂than␈α∂kept␈α∂in␈α∂core.␈α∞ While
␈↓ α←␈↓this␈α∂is␈α∞accomplished␈α∂with␈α∞a␈α∂very␈α∞efficient␈α∂hashing␈α∞routine␈↓
3␈↓␈α∂the␈α∂program␈α∞is
␈↓ α←␈↓already␈αtoo␈αslow␈αwith␈αnormal␈αmachine␈αloading--it␈αis␈αcaught␈αon␈αboth␈αends␈αof
␈↓ α←␈↓the␈α_speed-space␈α↔trade-off.␈α_ Possible␈α↔solutions␈α_lie␈α↔both␈α_in␈α_increases␈α↔in
␈↓ α←␈↓hardware resources and the use of faster programming languages.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α∀this␈α∀report␈α∀started␈α∀with␈α∀the␈α∀suggestion␈α∀that␈α∀it␈α∀would␈α∀be
␈↓ α←␈↓advantageous␈αto␈αreplace␈α
the␈αassistant␈αindicated␈α
in␈αFig.␈α1-1␈α
by␈αa␈αprogram␈α
that
␈↓ α←␈↓allows␈α
the␈α∞expert␈α
to␈α∞educate␈α
the␈α
performance␈α∞program␈α
directly.␈α∞ This␈α
offers
␈↓ α←␈↓many␈α∞advantages␈α∞from␈α∞the␈α∞point␈α∞of␈α∞view␈α∞of␈α∞speed,␈α∞manpower,␈α∞and␈α∂ease␈α∞of
␈↓ α←␈↓knowledge␈αbase␈αconstruction.␈α But␈αwhat␈αkinds␈αof␈αproblems␈αmight␈αit␈αproduce?
␈↓ α←␈↓Experience␈α∀has␈α∪demonstrated␈α∀that␈α∪a␈α∀certain␈α∪period␈α∀of␈α∀acclimatization␈α∪is
␈↓ α←␈↓necessary␈α∞before␈α∞new␈α∞experts␈α∞are␈α∞familiar␈α∞enough␈α∞with␈α∂the␈α∞␈↓↓weltanschauung␈↓
␈↓ α←␈↓implicit␈α
in␈α
the␈α
performance␈α
program␈α
to␈α
be␈α
able␈α
to␈α
view␈α
the␈α
domain␈α
and␈αits
␈↓ α←␈↓challenges␈α∪in␈α∩those␈α∪terms.␈α∩ During␈α∪that␈α∩period,␈α∪the␈α∩task␈α∪of␈α∪the␈α∩assistant
␈↓ α←␈↓becomes␈α∩more␈α∪than␈α∩simply␈α∪translating␈α∩the␈α∪clinician's␈α∩statements␈α∪into␈α∩␈↓¬LISP␈↓
␈↓ α←␈↓rules;␈α
it␈α
includes␈α
re-interpreting␈α
them␈α
as␈α
well.␈α
 If␈α
the␈α
expert␈α
is␈α
put␈α
in␈αdirect
␈↓ α←␈↓contact␈α⊂with␈α⊂the␈α∂program,␈α⊂what␈α⊂problems␈α∂might␈α⊂arise␈α⊂from␈α∂(unrecognized)

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[2]␈αOr␈αa␈αset␈αof␈αrules,␈αplus␈αa␈αmeta-rule␈αfor␈αordering.␈α This␈αapproach␈αassumes
␈↓ α←␈↓that,␈α⊃when␈α⊂feasible,␈α⊃it␈α⊂is␈α⊃more␈α⊂profitable␈α⊃in␈α⊂the␈α⊃long␈α⊂run␈α⊃to␈α⊂use␈α⊃a␈α⊂single
␈↓ α←␈↓representation␈α_(in␈α_this␈α→case,␈α_rules)␈α_and␈α→provide␈α_ways␈α_to␈α→translate␈α_all
␈↓ α←␈↓knowledge␈α⊗into␈α⊗this␈α⊗form,␈α∃than␈α⊗to␈α⊗introduce␈α⊗new␈α⊗representations,␈α∃each
␈↓ α←␈↓tailored␈αto␈α
a␈αspecific␈αtask.␈α
Experience␈αwith␈α␈↓¬TEIRESIAS␈↓␈α
so␈αfar␈αsuggests␈α
that␈αthis
␈↓ α←␈↓is true, but the issue is still far from settled.

␈↓"β␈↓ α←␈↓[3]␈α
The␈α
routine␈αwas␈α
provided␈α
by␈αLarry␈α
Masinter␈α
of␈αXerox␈α
PARC␈α
and␈αBill
␈↓ α←␈↓van Melle of the ␈↓¬MYCIN␈↓ group.
␈↓ α←␈↓␈↓8-3␈↓ λGLOBAL LIMITATIONS    261␈↓

␈↓"β␈↓ α←␈↓differences␈α∞in␈α∞their␈α∞models␈α∞of␈α∞the␈α
world?␈α∞ How␈α∞much␈α∞of␈α∞what␈α∞the␈α
assistant
␈↓ α←␈↓typically␈αdoes␈αin␈αmediating␈αbetween␈αthe␈αexpert␈αand␈αsystem␈αinvolves␈αa␈αshift␈α
in
␈↓ α←␈↓representations␈α∀and␈α∃a␈α∀resolution␈α∃of␈α∀such␈α∀conflicts?␈α∃ Can␈α∀this␈α∃ability␈α∀be
␈↓ α←␈↓formalized␈α
and␈α
made␈α
a␈α∞part␈α
of␈α
the␈α
system␈α∞itself?␈α
 It␈α
appears␈α
to␈α∞be␈α
difficult,
␈↓ α←␈↓requiring,␈α∂as␈α∂it␈α∂does,␈α∂solutions␈α∂to␈α∞problems␈α∂known␈α∂to␈α∂be␈α∂difficult,␈α∂and␈α∞will
␈↓ α←␈↓require a good deal of further work.
␈↓ α←␈↓␈↓262    CONCLUSIONS␈↓ 
#8-4␈↓

␈↓"β␈↓ α←␈↓␈↓α8-4    THE OTHER THEMES; SOME SPECULATIONS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Here␈αwe␈αreturn␈αat␈αlast␈αto␈αthe␈αlist␈αof␈αthemes␈αgiven␈αin␈αchapter␈α1.␈α Each
␈↓ α←␈↓is␈α
described␈α
in␈α
terms␈α
of␈α∞three␈α
topics: ␈α
(␈↓↓i␈↓)␈α
meaning: ␈α
The␈α∞substantive␈α
concept
␈↓ α←␈↓behind␈αthe␈αcatchphrase␈αis␈αexplained␈αand␈αits␈αutility␈αmade␈αclear;␈α(␈↓↓ii␈↓)␈αexamples: 
␈↓ α←␈↓Several␈αexamples␈αare␈αgiven␈αof␈αhow␈αthe␈αidea␈αhas␈αbeen␈αemployed␈α
in␈α␈↓¬TEIRESIAS␈↓;
␈↓ α←␈↓and␈α
(␈↓↓iii␈↓)␈α
evaluation␈α
and␈α
prognosis:␈α∞The␈α
examples␈α
are␈α
evaluated␈α
in␈α∞terms␈α
of
␈↓ α←␈↓the␈α∩goals␈α∩suggested␈α∪by␈α∩the␈α∩phrases,␈α∩and␈α∪an␈α∩indication␈α∩is␈α∩given␈α∪of␈α∩what
␈↓ α←␈↓difficult problems remain.
␈↓"β␈↓ α←␈↓␈↓ β?Some␈α∞warnings␈α∞first: ␈α
Despite␈α∞attempts␈α∞to␈α
linearize␈α∞this␈α∞discussion,␈α
it
␈↓ α←␈↓remains␈α≥a␈α≡network␈α≥of␈α≡interrelated␈α≥ideas.␈α≡ The␈α≥network␈α≡is␈α≥densely
␈↓ α←␈↓interconnected,␈α∩and␈α∩most␈α∩of␈α∩the␈α∩ideas␈α⊃will␈α∩lead␈α∩to␈α∩most␈α∩of␈α∩the␈α∩others␈α⊃if
␈↓ α←␈↓followed␈αfar␈αenough.␈α To␈αmake␈αthe␈αdiscussion␈αcomprehensible,␈αmany␈αof␈α
those
␈↓ α←␈↓connections␈α∞have␈α∞been␈α∂ignored,␈α∞except␈α∞where␈α∂truly␈α∞necessary.␈α∞ Even␈α∂so,␈α∞the
␈↓ α←␈↓discussion occasionally loops back on itself.
␈↓"β␈↓ α←␈↓␈↓ β?It␈α∂also␈α∂tends␈α∂to␈α∂raise␈α∂a␈α∂good␈α∂many␈α∂more␈α∂questions␈α∂than␈α∂it␈α∞answers.
␈↓ α←␈↓After␈αpointing␈αout␈αhow␈αthese␈αideas␈αhave␈αbeen␈αused␈αthroughout␈αthe␈αprogram,
␈↓ α←␈↓it␈α∞looks␈α∞forward␈α∞and␈α∞speculates␈α∞about␈α∞where␈α∞some␈α∞of␈α∞this␈α∂might␈α∞eventually
␈↓ α←␈↓lead.

␈↓"β␈↓ α←␈↓␈↓α8-4-1    Why write a program:  Two views␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈α⊃first␈α⊂five␈α⊃themes␈α⊂are␈α⊃grouped␈α⊂together␈α⊃because␈α⊃they␈α⊂represent
␈↓ α←␈↓different␈α∞techniques␈α∞involved␈α∞in␈α∞a␈α∞certain␈α∞style␈α∞of␈α∞programming.␈α∂ To␈α∞make
␈↓ α←␈↓that␈α∃style␈α⊗clear,␈α∃consider␈α⊗two␈α∃different,␈α⊗very␈α∃broad␈α⊗characterizations␈α∃of
␈↓ α←␈↓situations in which computer programs might be written.
␈↓"β␈↓ α←␈↓␈↓ β?Traditionally,␈α
programs␈α
have␈α
been␈αwritten␈α
to␈α
solve␈α
specific␈αproblems
␈↓ α←␈↓and␈α∪obtain␈α∩``an␈α∪answer.'' ␈α∪The␈α∩answer␈α∪may␈α∪be␈α∩of␈α∪many␈α∪types,␈α∩including
␈↓ α←␈↓numeric␈α≡(as␈α≥in␈α≡business␈α≡and␈α≥scientific␈α≡computation),␈α≡symbolic␈α≥(e.g.,
␈↓ α←␈↓information␈α⊂retrieval),␈α⊂or␈α⊂others␈α⊂of␈α∂a␈α⊂more␈α⊂abstract␈α⊂form.␈α⊂The␈α⊂main␈α∂issue
␈↓ α←␈↓here␈α⊗is␈α⊗that␈α↔these␈α⊗programs␈α⊗are␈α⊗applied␈α↔to␈α⊗tasks␈α⊗with␈α↔two␈α⊗important
␈↓ α←␈↓characteristics: ␈α∀(1)  There␈α∀is␈α∀``an␈α∃answer''␈α∀to␈α∀the␈α∀problem,␈α∀and␈α∃(2)  it␈α∀is
␈↓ α←␈↓possible␈α
to␈α∞speak␈α
of␈α∞a␈α
``final,''␈α∞debugged␈α
version␈α
of␈α∞a␈α
program␈α∞that␈α
is␈α∞to␈α
be
␈↓ α←␈↓used␈α∞extensively␈α
as␈α∞it␈α
is,␈α∞with␈α∞relatively␈α
long␈α∞periods␈α
between␈α∞changes␈α∞to␈α
it.
␈↓ α←␈↓This␈α⊃view␈α⊃encourages␈α∩a␈α⊃style␈α⊃of␈α∩programming␈α⊃in␈α⊃which␈α∩the␈α⊃programmer
␈↓ α←␈↓spends␈α
a␈αlong␈α
time␈α
thinking␈αabout␈α
the␈α
problem␈αfirst,␈α
tries␈α
to␈αsolve␈α
as␈αmuch␈α
of
␈↓ α←␈↓it␈α
as␈α
possible␈αby␈α
hand,␈α
and␈αthen␈α
abstracts␈α
out␈αonly␈α
the␈α
very␈α
end-product␈αof
␈↓ α←␈↓all␈α
that␈α
thought␈α
to␈α
be␈α
embodied␈α
in␈α
the␈α
program.␈α
 That␈α
is,␈α
the␈α
program␈α
can
␈↓ α←␈↓become␈αsimply␈αa␈αway␈αof␈αmanipulating␈αsymbols␈αto␈αprovide␈α``the␈αanswer,''␈αwith
␈↓ α←␈↓little␈α
indication␈α∞of␈α
what␈α∞the␈α
original␈α∞problem␈α
was␈α∞or,␈α
more␈α∞important,␈α
what
␈↓ α←␈↓knowledge was required to solve it.
␈↓"β␈↓ α←␈↓␈↓ β?Large,␈α⊂knowledge-based␈α⊂programs␈α⊂appear␈α⊂to␈α⊂have␈α⊂a␈α⊂fundamentally
␈↓ α←␈↓different␈α⊃character.␈α⊂ As␈α⊃noted␈α⊂earlier,␈α⊃the␈α⊂construction␈α⊃of␈α⊃large␈α⊂knowledge
␈↓ α←␈↓bases␈αis␈αa␈α
long-term␈αoperation␈αthat␈α
is␈αnever␈αreally␈α
``finished.'' ␈αNot␈αonly␈αis␈α
the
␈↓ α←␈↓approach␈α⊂to␈α⊂competence␈α⊂incremental␈α⊂(and␈α⊂occasionally␈α⊂asymptotic),␈α⊂but␈α∂the
␈↓ α←␈↓fields␈αto␈αwhich␈α
such␈αprograms␈αare␈αapplied␈α
are␈αtypically␈αthose␈αwhich␈α
are␈αstill
␈↓ α←␈↓under␈αactive␈α
development.␈α The␈α
knowledge␈αbase␈α
is␈αthus␈α
inherently␈αa␈α
dynamic
␈↓ α←␈↓␈↓8-4␈↓ ¬yTHE OTHER THEMES; SOME SPECULATIONS    263␈↓

␈↓"β␈↓ α←␈↓structure.␈α The␈αaim␈αhere␈αis␈αthus␈αnot␈αsimply␈αto␈αbuild␈αa␈αprogram␈αthat␈αexhibits
␈↓ α←␈↓a␈α⊃certain␈α⊃specified␈α⊃behavior,␈↓
4␈↓␈α⊃but␈α⊃␈↓↓to␈α⊃use␈α⊃the␈α⊃program␈α⊃construction␈α⊂process
␈↓ α←␈↓↓itself␈α
as␈αa␈α
way␈αof␈α
explicating␈αknowledge␈α
in␈αthe␈α
field,␈αand␈α
to␈αuse␈α
the␈αprogram
␈↓ α←␈↓↓text␈α
as␈αa␈α
medium␈αof␈α
expression␈αof␈α
many␈α
forms␈αof␈α
knowledge␈αabout␈α
the␈αtask␈α
and
␈↓ α←␈↓↓its␈αsolution␈↓.␈α That␈αis,␈αthe␈α
program␈αbecomes␈αmore␈αthan␈αa␈αcollection␈α
of␈αsymbol
␈↓ α←␈↓manipulation␈α∞instructions.␈α∂ It␈α∞is␈α∂used␈α∞as␈α∂an␈α∞environment␈α∂in␈α∞which␈α∂to␈α∞elicit,
␈↓ α←␈↓collect, and store large amounts of diverse forms of knowledge.
␈↓"β␈↓ α←␈↓␈↓ β?To␈αsum␈αup␈αthese␈αtwo␈αlong␈αstatements␈αin␈αa␈αsingle␈αphrase: ␈αIn␈αone␈αcase
␈↓ α←␈↓it␈α∂is␈α∞reasonable␈α∂to␈α∂think␈α∞in␈α∂terms␈α∞of␈α∂a␈α∂program␈α∞that␈α∂represents␈α∂a␈α∞complete
␈↓ α←␈↓solution␈αto␈αa␈α
problem,␈αand␈αhence␈αencode␈α
only␈αthe␈αminimally␈αnecessary␈α
symbol
␈↓ α←␈↓manipulation␈αsteps␈α
in␈αthat␈α
solution;␈αin␈αthe␈α
other,␈αthe␈α
process␈αis␈α
a␈αcontinuing
␈↓ α←␈↓one of codifying and accumulating information.
␈↓"β␈↓ α←␈↓␈↓ β?This␈αsecond␈αview␈αhas␈α
been␈αadopted␈αin␈αbuilding␈αthe␈α
knowledge␈αbases
␈↓ α←␈↓needed␈α∂for␈α∂several␈α∂performance␈α∂programs␈α⊂over␈α∂the␈α∂years␈α∂and␈α∂has␈α⊂made␈α∂it
␈↓ α←␈↓possible␈αto␈αorganize␈αand␈αrepresent␈αthe␈αlarge␈αamounts␈αof␈αknowledge␈αrequired.
␈↓ α←␈↓Question: ␈α⊂Is␈α⊂it␈α⊂possible␈α⊂to␈α⊂turn␈α⊂the␈α⊂idea␈α⊂in␈α⊂on␈α⊂itself␈α⊂and␈α⊂apply␈α⊂it␈α⊃to␈α⊂the
␈↓ α←␈↓internal␈α∂world,␈α∂the␈α⊂world␈α∂of␈α∂representations␈α⊂and␈α∂data␈α∂structures␈α⊂inside␈α∂the
␈↓ α←␈↓program? ␈αWill␈αit␈α
be␈αas␈αuseful␈α
at␈αthe␈αmeta-level␈α
as␈αit␈αwas␈α
at␈αthe␈αobject␈α
level? 
␈↓ α←␈↓Can␈α∃it␈α∃help␈α∃to␈α∃assemble,␈α⊗organize,␈α∃and␈α∃maintain␈α∃the␈α∃large␈α⊗amount␈α∃of
␈↓ α←␈↓knowledge␈α⊃about␈α⊃representations␈α⊃and␈α⊃data␈α⊃structures␈α⊃that␈α⊃is␈α⊃necessary␈α⊂for
␈↓ α←␈↓knowledge␈α⊂base␈α⊂construction␈α⊂and␈α∂maintenance?␈α⊂ Other␈α⊂efforts␈α⊂in␈α⊂this␈α∂vein
␈↓ α←␈↓have␈α∩centered␈α∩around␈α∪the␈α∩use␈α∩of␈α∪more␈α∩formal␈α∩methods␈α∪(e.g.,␈α∩[Suzuki76],
␈↓ α←␈↓[Spitzen75]).␈α∂ Will␈α∞the␈α∂essentially␈α∞informal,␈α∂heuristic␈α∞techniques␈α∂we␈α∂want␈α∞to
␈↓ α←␈↓employ offer sufficient power?
␈↓"β␈↓ α←␈↓␈↓ β?The␈αattempt␈α
to␈αanswer␈α
these␈αquestions␈αaffirmatively␈α
has␈αbeen␈α
one␈αof
␈↓ α←␈↓the central themes of this work.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αimmediate␈αapplication␈αof␈αthese␈αideas␈αhas␈αbeen␈αto␈αprovide␈αa␈αbasis
␈↓ α←␈↓for␈α⊂the␈α⊂tools␈α⊂described␈α⊂in␈α⊂earlier␈α⊂chapters.␈α⊂ More␈α⊃fundamentally,␈α⊂however,
␈↓ α←␈↓they␈αoffer␈αthe␈αpossibility␈αof␈αgetting␈αa␈αstep␈αcloser␈αto␈αthat␈αseductive␈αpromise␈αof
␈↓ α←␈↓the␈α∩stored␈α∪program␈α∩computer:␈α∪introspection.␈α∩ As␈α∩we␈α∪have␈α∩seen,␈α∪the␈α∩most
␈↓ α←␈↓useful␈αforms␈αof␈αmeta-level␈αknowledge␈αare␈αthose␈αthat␈αmake␈αit␈αpossible␈αfor␈αthe
␈↓ α←␈↓system␈α⊃to␈α⊃examine␈α⊃its␈α⊃own␈α⊂knowledge␈α⊃directly.␈α⊃ Every␈α⊃step␈α⊃removed␈α⊂from
␈↓ α←␈↓direct␈α∃examination␈α∃represents␈α⊗a␈α∃level␈α∃of␈α∃defeat␈α⊗and␈α∃loses␈α∃some␈α⊗of␈α∃the
␈↓ α←␈↓advantages.␈α Thus␈α
the␈αrule␈αmodels␈α
are␈αderived␈αfrom␈α
rules␈αdirectly␈αand␈α
adjust
␈↓ α←␈↓to␈α⊂changes␈α⊂in␈α⊃the␈α⊂knowledge␈α⊂base,␈α⊃while␈α⊂the␈α⊂explanation␈α⊃routines␈α⊂require
␈↓ α←␈↓rewriting␈α
whenever␈α
the␈α
control␈α
structure␈α
changes,␈α
because␈α
they␈α
do␈αnot␈α
directly
␈↓ α←␈↓reference the code of the system they are trying to explain.
␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[4]␈α∩Indeed,␈α∩it␈α⊃is␈α∩not␈α∩even␈α⊃clear␈α∩whether␈α∩a␈α⊃behavioral␈α∩definition␈α∩of␈α⊃such
␈↓ α←␈↓programs␈α∪makes␈α∪sense.␈α∪ In␈α∪other␈α∪cases␈α∪behavioral␈α∪definitions␈α∪have␈α∪been
␈↓ α←␈↓proposed␈α
and␈αused␈α
(e.g.,␈αfor␈α
automatic␈αprogramming␈α
as␈αin␈α
[Shaw75]).␈α They
␈↓ α←␈↓make␈αsense␈α
in␈αthat␈αcase␈α
because␈αthey␈α
are␈αa␈αmore␈α
compact␈αexpression␈αthan␈α
the
␈↓ α←␈↓program␈α
that␈αproduces␈α
the␈αbehavior.␈α
 It␈αis␈α
not␈αclear␈α
that␈αthis␈α
would␈α
be␈αtrue
␈↓ α←␈↓of␈α∞large␈α∞knowledge␈α
based␈α∞programs--it␈α∞appears␈α
difficult␈α∞to␈α∞give␈α∞a␈α
complete
␈↓ α←␈↓specification␈α∀for␈α∀their␈α∀behavior␈α∀in␈α∪any␈α∀terms␈α∀that␈α∀produce␈α∀a␈α∪definition
␈↓ α←␈↓smaller than the actual knowledge base.
␈↓ α←␈↓␈↓264    CONCLUSIONS␈↓ 
#8-4␈↓

␈↓"β␈↓ α←␈↓␈↓ β?With␈α
this␈α∞overview,␈α
let's␈α∞consider␈α
some␈α∞of␈α
the␈α∞specifics␈α
and␈α∞see␈α
how
␈↓ α←␈↓the first five themes suggest steps in the direction of these goals.

␈↓ α←␈↓␈↓αTheme 1: Task-specific high-level languages make code easier to read.␈↓
␈↓"β␈↓ α←␈↓␈↓ β?``Task-specific␈α_high-level␈α_language''␈α_refers,␈α_simply␈α_enough,␈α_to␈α_a
␈↓ α←␈↓language␈α
in␈αwhich␈α
the␈α
conceptual␈αprimitives␈α
are␈αtask␈α
specific.␈α
 The␈αprimary
␈↓ α←␈↓example␈αof␈αthis␈αthat␈αwe␈αhave␈αseen␈αis␈αthe␈αset␈αof␈αrules,␈αcomposed␈αof␈αprimitives
␈↓ α←␈↓like␈α∂␈↓↓attribute␈↓␈α∞and␈α∂␈↓↓value␈↓,␈α∂whose␈α∞instances␈α∂in␈α∂␈↓¬MYCIN␈↓␈α∞include␈α∂culture␈α∂␈↓∧SITE␈↓␈α∞and
␈↓ α←␈↓␈↓∧BLOOD␈↓.␈α∞ The␈α∞rules␈α∞were␈α∞discussed␈α∞in␈α∞these␈α∞terms␈α∞in␈α∞Section␈α∞2-4-4,␈α∂where␈α∞it
␈↓ α←␈↓was␈αnoted␈αthat␈αthey␈αare␈αin␈αeffect␈αa␈αdomain-specific␈αhigh-level␈αlanguage␈αwith
␈↓ α←␈↓a rule as the sole statement type.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∃are␈α∃several␈α∃other␈α∃examples,␈α∃many␈α∃of␈α∃them␈α∃found␈α∃in␈α∃the
␈↓ α←␈↓schemata␈α∃described␈α∀in␈α∃chapter␈α∃6.␈α∀ Included␈α∃are␈α∀the␈α∃``language''␈α∃of␈α∀data
␈↓ α←␈↓structure␈α↔␈↓∧RELATIONS␈↓,␈α⊗the␈α↔set␈α⊗of␈α↔␈↓↓slotnames␈↓␈α⊗of␈α↔which␈α⊗the␈α↔schemata␈α⊗are
␈↓ α←␈↓composed, and the several types of ␈↓↓advice␈↓ found in slots.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αforemost␈αreason␈αfor␈αdesigning␈α
and␈αusing␈α``languages''␈αlike␈αthese␈α
is
␈↓ α←␈↓to␈α∀make␈α∪possible␈α∀what␈α∪we␈α∀might␈α∪label␈α∀``top-down␈α∀code␈α∪understanding.'' 
␈↓ α←␈↓Previous␈α∃efforts␈α∃at␈α∃building␈α∃program␈α∃understanding␈α∃systems␈α∃have␈α∀been
␈↓ α←␈↓aimed␈α≥at␈α≤several␈α≥goals,␈α≤including␈α≥proving␈α≤programs␈α≥correct␈α≥(as␈α≤in
␈↓ α←␈↓[Waldinger74]␈αand␈α
[Manna69])␈αand␈α
for␈αuse␈α
in␈αautomatic␈α
programming␈α(as␈α
in
␈↓ α←␈↓[Green74]).␈α∞Most␈α
of␈α∞these␈α
systems␈α∞attempt␈α
to␈α∞assign␈α
meaning␈α∞to␈α
the␈α∞code␈α
of
␈↓ α←␈↓some␈α∩standard␈α∪(domain␈α∩independent)␈α∩programming␈α∪language␈α∩like␈α∪␈↓¬LISP␈↓␈α∩or
␈↓ α←␈↓␈↓¬ALGOL␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈α∃the␈α∀problems␈α∃encountered␈α∃in␈α∀doing␈α∃this␈α∀are␈α∃known␈α∃to␈α∀be
␈↓ α←␈↓difficult,␈α
we␈α∞have␈α
used␈α∞the␈α
high-level␈α
``languages''␈α∞as␈α
a␈α∞convenient␈α
shortcut.
␈↓ α←␈↓Rather␈α∀than␈α∀attempting␈α∀to␈α∃assign␈α∀formal␈α∀semantics␈α∀to␈α∀ordinary␈α∃code,␈α∀a
␈↓ α←␈↓``meaning''␈α⊗is␈α⊗assigned␈α⊗to␈α⊗each␈α∃of␈α⊗the␈α⊗primitives␈α⊗in␈α⊗the␈α⊗language␈α∃and
␈↓ α←␈↓represented␈α
in␈αone␈α
or␈α
more␈αinformal␈α
ways.␈α
 For␈αexample,␈α
part␈α
of␈αthe␈α
meaning
␈↓ α←␈↓of␈αthe␈α
concept␈α␈↓↓attribute␈↓␈α
is␈αrepresented␈α
by␈αthe␈α
routine␈αassociated␈α
with␈αit␈αthat␈α
is
␈↓ α←␈↓invoked␈α⊂during␈α⊂new␈α⊂rule␈α⊂acquisition␈α⊂(see␈α⊂Section␈α⊂5-4-2).␈α⊂ Another␈α⊂part␈α∂is
␈↓ α←␈↓represented␈αby␈αthe␈αcode␈αin␈αthe␈αexplanation␈αroutines␈αthat␈αchecks␈αthe␈αattribute
␈↓ α←␈↓in␈αa␈αpremise␈αclause␈αwhen␈αdividing␈αa␈αrule␈αpremise␈αinto␈αknown␈αand␈αunknown
␈↓ α←␈↓clauses.␈α∂ The␈α⊂inequality␈α∂described␈α⊂in␈α∂Section␈α⊂3-7␈α∂is␈α⊂another␈α∂example;␈α⊂it␈α∂is
␈↓ α←␈↓part of the ``meaning'' of a predicate function.
␈↓"β␈↓ α←␈↓␈↓ β?Since␈α↔this␈α↔technique␈α↔has␈α↔been␈α↔employed␈α↔extensively␈α↔here,␈α↔it␈α⊗is
␈↓ α←␈↓important␈α≠to␈α≠consider␈α~its␈α≠limitations.␈α≠ While␈α~it␈α≠does␈α≠make␈α~program
␈↓ α←␈↓understanding␈α⊃easier,␈α⊂it␈α⊃approaches␈α⊂the␈α⊃task␈α⊂at␈α⊃a␈α⊂higher␈α⊃conceptual␈α⊂level,
␈↓ α←␈↓which␈α∃makes␈α∃the␈α∃result␈α⊗correspondingly␈α∃less␈α∃powerful.␈α∃ We␈α⊗cannot,␈α∃for
␈↓ α←␈↓instance,␈αprove␈αthat␈αthe␈αimplementation␈αof␈α␈↓∧SAME␈↓␈αis␈αcorrect,␈αbut␈αwe␈αcan␈αuse␈αits
␈↓ α←␈↓indicated␈α
``meaning''␈α
in␈αthe␈α
ways␈α
illustrated.␈α More␈α
fundamentally,␈α
the␈αentire
␈↓ α←␈↓approach␈α≠depends␈α≠on␈α≠the␈α≠existence␈α≠of␈α≠a␈α≠finite␈α≠number␈α≠of␈α≠``mostly
␈↓ α←␈↓independent''␈α∂primitives.␈α⊂ This␈α∂means␈α∂a␈α⊂set␈α∂of␈α∂primitives␈α⊂with␈α∂only␈α⊂a␈α∂few,
␈↓ α←␈↓well-specified␈αinteractions␈αbetween␈αthem.␈α The␈αnumber␈αof␈αinteractions␈α
should
␈↓ α←␈↓be␈α⊃far␈α⊃less␈α⊃than␈α⊃the␈α⊃total␈α⊃that␈α⊃is␈α⊃possible␈α⊃and␈α⊃interactions␈α⊃that␈α∩do␈α⊃occur
␈↓ α←␈↓should␈αbe␈α
uncomplicated.␈α In␈αthe␈α
case␈αof␈αthe␈α
rule␈αlanguage,␈αfor␈α
instance,␈αthe
␈↓ α←␈↓␈↓8-4␈↓ ¬yTHE OTHER THEMES; SOME SPECULATIONS    265␈↓

␈↓"β␈↓ α←␈↓concepts␈α∂of␈α⊂certainty␈α∂factor␈α∂and␈α⊂attribute␈α∂are␈α∂independent,␈α⊂while␈α∂attributes
␈↓ α←␈↓and␈α∃values␈α∃have␈α∃a␈α⊗well-specified␈α∃relation␈α∃that␈α∃is␈α∃reasonably␈α⊗simple␈α∃to
␈↓ α←␈↓represent.
␈↓"β␈↓ α←␈↓␈↓ β?This␈αis␈α
an␈αimportant␈αlimitation,␈α
one␈αthat␈α
we␈αhave␈αencountered␈α
earlier.
␈↓ α←␈↓Recall␈α∂the␈α∂discussion␈α∂in␈α∂Section␈α∞3-9␈α∂on␈α∂the␈α∂deficiencies␈α∂of␈α∂the␈α∞explanation
␈↓ α←␈↓routines␈α⊃that␈α⊃arise␈α∩from␈α⊃the␈α⊃inability␈α∩to␈α⊃represent␈α⊃control␈α∩structures.␈α⊃ We
␈↓ α←␈↓pointed␈α∩out␈α∪there␈α∩the␈α∩utility␈α∪of␈α∩a␈α∩language␈α∪of␈α∩``intentions''␈α∩in␈α∪making␈α∩it
␈↓ α←␈↓possible for the system to explain its actions, yet were unable to propose one.
␈↓"β␈↓ α←␈↓␈↓ β?When␈α∩it␈α∩is␈α∩possible␈α∩to␈α∩find␈α∩such␈α∩a␈α∩language,␈α∩however,␈α∩it␈α∩can␈α⊃be
␈↓ α←␈↓applied␈αin␈αa␈αvariety␈αof␈αways.␈α
 In␈αthe␈αsimplest␈αcase,␈αthe␈α``meaning''␈αof␈α
each␈αof
␈↓ α←␈↓the␈α→primitives␈α→is␈α_considered␈α→and␈α→used␈α_individually.␈α→ A␈α→slightly␈α_more
␈↓ α←␈↓sophisticated␈α→use␈α~allows␈α→combinations␈α→of␈α~primitives␈α→to␈α~describe␈α→other
␈↓ α←␈↓structures.␈α_ Consider,␈α→for␈α_example,␈α_the␈α→templates␈α_associated␈α→with␈α_each
␈↓ α←␈↓predicate␈α∞function.␈α∂ The␈α∞ability␈α∂to␈α∞dissect␈α∂a␈α∞given␈α∂function␈α∞call␈α∂could␈α∞have
␈↓ α←␈↓been␈α⊗based␈α⊗on␈α↔individual,␈α⊗hand-tailored␈α⊗routines␈α⊗associated␈α↔with␈α⊗each
␈↓ α←␈↓function.␈α∞ Instead,␈α∞a␈α∞lower␈α∞level␈α∞of␈α
detail␈α∞was␈α∞used.␈α∞ This␈α∞makes␈α∞it␈α
possible
␈↓ α←␈↓for␈αeach␈α
function␈αto␈α
carry␈αinformation␈αdescribing␈α
its␈αown␈α
calls␈αand␈αmakes␈α
the
␈↓ α←␈↓system␈α∂extensible: ␈α∂As␈α∂long␈α∂as␈α∂calls␈α∂to␈α∂a␈α∂new␈α∂function␈α∂can␈α∂be␈α∂described␈α∂in
␈↓ α←␈↓terms␈α∂of␈α∞the␈α∂available␈α∞primitives,␈α∂the␈α∞function␈α∂can␈α∞be␈α∂added␈α∞to␈α∂the␈α∞system
␈↓ α←␈↓without any change to existing facilities.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α
idea␈αof␈α
``top-down␈α
code␈αunderstanding''␈α
is␈α
the␈αprimary␈α
technique
␈↓ α←␈↓we␈αhave␈αused␈αto␈αprovide␈αa␈αdegree␈αof␈αintrospection.␈α The␈αsystem␈αcan␈αexamine
␈↓ α←␈↓parts␈α∀of␈α∀its␈α∀own␈α∀structure␈α∀and␈α∀can␈α∀``understand''␈α∀them␈α∀in␈α∀terms␈α∀of␈α∪the
␈↓ α←␈↓primitives of the high-level languages used.

␈↓ α←␈↓␈↓αTheme 2:  Knowledge in programs should be explicit and accessible.␈↓
␈↓"β␈↓ α←␈↓␈↓ β?Consider the following two lists, taken from the current ␈↓¬MYCIN␈↓ system:

␈↓"β␈↓ α←␈↓	(BLOOD BONE BRAIN CSF JOINT LIVER LUNG MUSCLE TENDON-SHEATH)
␈↓"β␈↓ α←␈↓	(RULE012 RULE146 RULE087 RULE234 RULE043 RULE101)

␈↓"β␈↓ α←␈↓␈↓ β?The␈α
first␈αappears␈α
to␈αbe␈α
ordered␈αalphabetically;␈α
the␈αordering,␈α
if␈αany,␈α
of
␈↓ α←␈↓the␈α∪second␈α∪is␈α∪unclear.␈α∪ To␈α∪avoid␈α∪introducing␈α∪bugs␈α∪when␈α∪adding␈α∪a␈α∩new
␈↓ α←␈↓element␈αto␈αeither␈αof␈αthem,␈αtwo␈αthings␈αare␈αimportant:␈α(a)  If␈αthe␈αlist␈αis␈αordered,
␈↓ α←␈↓what␈α∩is␈α∪the␈α∩ordering␈α∪criterion?␈α∩and␈α∪(b)  What␈α∩is␈α∪the␈α∩significance␈α∪of␈α∩the
␈↓ α←␈↓ordering (i.e., who depends on it)?
␈↓"β␈↓ α←␈↓␈↓ β?The␈α∂answer␈α⊂to␈α∂the␈α∂first␈α⊂question␈α∂is␈α∂required␈α⊂in␈α∂order␈α⊂to␈α∂determine
␈↓ α←␈↓where␈αin␈αthe␈αlist␈αthe␈αnew␈αelement␈αis␈αto␈αbe␈αplaced.␈αThe␈αsecond␈αis␈αimportant␈αif
␈↓ α←␈↓we␈α
wish␈α
to␈α
consider␈α
violating␈α
the␈α
established␈α
order: ␈α
It␈α
gives␈α
a␈α
clear␈α
picture␈α
of
␈↓ α←␈↓the␈α∩consequences␈α∪and␈α∩indicates␈α∩where␈α∪changes␈α∩may␈α∩have␈α∪to␈α∩be␈α∪made␈α∩if
␈↓ α←␈↓serious problems result.
␈↓"β␈↓ α←␈↓␈↓ β?In␈αthe␈αcase␈αof␈αthe␈αfirst␈αlist,␈αthe␈αordering␈αis␈αin␈αfact␈αalphabetic,␈αbut␈αthis
␈↓ α←␈↓has␈αno␈αsignificance␈αwhatever␈αto␈αthe␈αrest␈αof␈αthe␈αsystem.␈α The␈αsecond␈αlist␈αhas␈αa
␈↓ α←␈↓partial␈α
ordering␈α
of␈α
its␈α
elements␈α
that␈α
is␈α
crucial␈α
to␈α
system␈α∞performance.␈α
 Until
␈↓ α←␈↓recently,␈αneither␈α
piece␈αof␈α
information␈αwas␈α
represented␈αanywhere␈α
in␈αthe␈α
system
␈↓ α←␈↓␈↓266    CONCLUSIONS␈↓ 
#8-4␈↓

␈↓"β␈↓ α←␈↓(nor␈αin␈α
fact␈αin␈α
any␈αdocumentation).␈α
The␈αsame␈α
phenomenon␈αmanifests␈αitself␈α
in
␈↓ α←␈↓many␈α∞forms␈α
in␈α∞almost␈α
all␈α∞large␈α
systems.␈α∞ There␈α
is␈α∞typically␈α
a␈α∞large␈α∞share␈α
of
␈↓ α←␈↓folklore␈α∞about␈α∞all␈α∂kinds␈α∞of␈α∞structures␈α∂that␈α∞resides␈α∞only␈α∂in␈α∞the␈α∞minds␈α∂of␈α∞the
␈↓ α←␈↓system␈α∪designers,␈α∩yet␈α∪such␈α∩information␈α∪is␈α∩often␈α∪crucial␈α∩to␈α∪the␈α∪process␈α∩of
␈↓ α←␈↓changing anything in the system.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α⊃is␈α⊃the␈α⊃kind␈α⊃of␈α⊃information␈α⊃we␈α⊃have␈α⊃tried␈α⊃to␈α⊃capture␈α∩with␈α⊃a
␈↓ α←␈↓variety␈α⊗of␈α⊗mechanisms.␈α⊗The␈α⊗first␈α⊗list␈α⊗above,␈α⊗for␈α⊗example,␈α⊗is␈α⊗classified
␈↓ α←␈↓internally␈α≠as␈α≤an␈α≠alphabetically␈α≠ordered␈α≤linear␈α≠list,␈α≠and␈α≤its␈α≠ordering
␈↓ α←␈↓information␈α
is␈αembodied␈α
in␈αthe␈α
updating␈α
function␈αassociated␈α
with␈αthe␈α
schema
␈↓ α←␈↓for␈α∞that␈α∂data␈α∞type.␈α∞ The␈α∂ordering␈α∞of␈α∂the␈α∞second␈α∞list␈α∂can␈α∞be␈α∂expressed␈α∞with
␈↓ α←␈↓meta-rules␈α
and,␈α
as␈α
noted␈αin␈α
Chapter␈α
5,␈α
this␈αmakes␈α
the␈α
problem␈α
of␈αupdating
␈↓ α←␈↓disappear.␈α The␈αlist␈αcan␈αbe␈αstored␈αunordered,␈αnew␈αrules␈αcan␈αsimply␈αbe␈αadded
␈↓ α←␈↓to the front, and the meta-rules will take care of reordering the list.
␈↓"β␈↓ α←␈↓␈↓ β?There␈αare␈αother␈αexamples␈αas␈αwell.␈α The␈αschema␈αnetwork,␈αfor␈αinstance,
␈↓ α←␈↓makes␈α∂explicit␈α∞the␈α∂interrelationships␈α∞of␈α∂the␈α∞data␈α∂types␈α∞in␈α∂the␈α∂system.␈α∞ Note
␈↓ α←␈↓that␈α⊂the␈α⊂point␈α⊂is␈α⊂not␈α⊃that␈α⊂this␈α⊂particular␈α⊂approach␈α⊂with␈α⊃its␈α⊂generalization
␈↓ α←␈↓hierarchy␈αis␈α
necessarily␈αthe␈α
way␈αthe␈αinformation␈α
should␈αbe␈α
represented.␈α We
␈↓ α←␈↓claim␈α⊂only␈α⊂that␈α⊂information␈α⊃about␈α⊂the␈α⊂internal␈α⊂organization␈α⊂of␈α⊃data␈α⊂types
␈↓ α←␈↓should␈α↔in␈α_some␈α↔fashion␈α↔be␈α_organized,␈α↔formalized,␈α↔and␈α_represented␈α↔in
␈↓ α←␈↓program-accessible terms.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
meta-rules␈α
offer␈αa␈α
more␈α
general␈α
example␈αof␈α
the␈α
theme.␈α
 One␈αof
␈↓ α←␈↓their␈α∂primary␈α∞contributions␈α∂was␈α∞to␈α∂provide␈α∞a␈α∂way␈α∞of␈α∂making␈α∂explicit␈α∞what
␈↓ α←␈↓most␈α∩programming␈α⊃languages␈α∩leave␈α∩implicit--the␈α⊃criteria␈α∩for␈α∩selecting␈α⊃the
␈↓ α←␈↓next knowledge source to be invoked.
␈↓"β␈↓ α←␈↓␈↓ β?A␈α⊂final␈α⊂example␈α∂worth␈α⊂recalling␈α⊂is␈α∂the␈α⊂restriction␈α⊂on␈α⊂rule␈α∂ordering
␈↓ α←␈↓required␈α∪for␈α∪the␈α∪``self-referencing''␈α∪rules␈α∪described␈α∪in␈α∪Section␈α∪7-4-4.␈α∩ To
␈↓ α←␈↓maintain␈αthe␈αcommutativity␈αof␈αcertainty␈αfactors,␈αall␈αnon-self-referencing␈α
rules
␈↓ α←␈↓must␈α
be␈αinvoked␈α
before␈α
any␈αthat␈α
are␈αself-referencing.␈α
 This␈α
important␈αpiece
␈↓ α←␈↓of␈α∂information␈α∂was␈α∂represented␈α∂first␈α∂in␈α∂an␈α∂obscure␈α∂piece␈α∂of␈α∂recursive␈α∂code
␈↓ α←␈↓buried␈α
in␈α
the␈α
control␈α
structure,␈α
later␈α
in␈α
a␈α
separate␈α
function,␈α
and␈α∞finally␈α
was
␈↓ α←␈↓captured␈α
by␈αa␈α
meta-rule␈αwhich␈α
expresses␈αthe␈α
necessary␈αordering␈α
in␈α
a␈αsimple
␈↓ α←␈↓and␈α⊗straightforward␈α⊗form.␈α⊗ It␈α⊗can␈α⊗be␈α⊗explained␈α⊗using␈α↔the␈α⊗explanation
␈↓ α←␈↓facilities␈αand␈α
changed,␈αif␈α
necessary,␈αusing␈αthe␈α
existing␈αrule␈α
editor.␈α It␈αhas␈α
thus
␈↓ α←␈↓progressed␈α∞from␈α∂being␈α∞implicit␈α∂and␈α∞inaccessible␈α∞to␈α∂being␈α∞more␈α∂explicit␈α∞and
␈↓ α←␈↓more easily retrieved.
␈↓"β␈↓ α←␈↓␈↓ β?Many␈α↔problems␈α↔remain,␈α⊗however.␈α↔ The␈α↔basic␈α⊗issue␈α↔here␈α↔is␈α⊗the
␈↓ α←␈↓standard␈α
difficult␈α
problem␈α
of␈α
representation␈α
of␈α
knowledge.␈α
 One␈α
indication␈α
of
␈↓ α←␈↓a␈α∂shortcoming␈α∂in␈α∞our␈α∂solution␈α∂is␈α∞hinted␈α∂at␈α∂by␈α∞the␈α∂comment␈α∂above␈α∂that␈α∞we
␈↓ α←␈↓have␈αused␈αa␈α``variety␈αof␈αmechanisms.'' ␈αOur␈αinformal␈αapproach␈αis␈αtoo␈αad␈αhoc
␈↓ α←␈↓and␈α∩needs␈α∪to␈α∩be␈α∩more␈α∪rigorous␈α∩before␈α∪it␈α∩can␈α∩claim␈α∪to␈α∩be␈α∪a␈α∩substantive
␈↓ α←␈↓solution.␈α↔ A␈α↔far␈α↔better␈α↔solution␈α↔would␈α↔provide␈α↔a␈α↔single␈α↔representation
␈↓ α←␈↓formalism␈αthat␈αcould␈αexpress␈αall␈αthe␈αdifferent␈αforms␈αof␈αinformation␈αrequired.
␈↓ α←␈↓Some␈α∩work␈α∩(e.g.,␈α∩[Suzuki76])␈α∩has␈α∩examined␈α∩the␈α∩applicability␈α∩of␈α⊃predicate
␈↓ α←␈↓calculus␈α∃to␈α∃similar␈α∀problems,␈α∃but␈α∃that␈α∀approach␈α∃has␈α∃not␈α∃yet␈α∀developed
␈↓ α←␈↓techniques␈α~capable␈α~of␈α~dealing␈α~with␈α~the␈α~size␈α~and␈α~complexity␈α~of␈α→the
␈↓ α←␈↓axiomatization and proof procedures required.
␈↓ α←␈↓␈↓8-4␈↓ ¬yTHE OTHER THEMES; SOME SPECULATIONS    267␈↓

␈↓"β␈↓ α←␈↓␈↓αTheme␈α3: ␈αPrograms␈αcan␈αbe␈αgiven␈αaccess␈αto␈αand␈αan␈αunderstanding␈αof␈αtheir
␈↓ α←␈↓αown representations.␈↓
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∩of␈α∩the␈α∩oldest␈α∩and␈α∩most␈α∩fundamental␈α∩issues␈α∩in␈α∩AI␈α∩is␈α∩that␈α∩of
␈↓ α←␈↓representation.␈α⊃ Different␈α⊃approaches␈α∩have␈α⊃come␈α⊃and␈α⊃gone␈α∩and␈α⊃generated
␈↓ α←␈↓innumerable␈α⊃discussions␈α⊂of␈α⊃respective␈α⊂power␈α⊃and␈α⊂virtue.␈α⊃ But␈α⊂in␈α⊃all␈α⊂these
␈↓ α←␈↓arguments,␈α∂one␈α∂entity␈α∂intimately␈α∂concerned␈α∂with␈α∂the␈α∂outcome␈α∂has␈α⊂been␈α∂left
␈↓ α←␈↓uninformed: the program itself.
␈↓"β␈↓ α←␈↓␈↓ β?Several␈α∀capabilities␈α∀follow␈α∀from␈α∀giving␈α∀a␈α∀program␈α∀the␈α∃ability␈α∀to
␈↓ α←␈↓examine␈αits␈αown␈αrepresentations.␈α The␈α
one␈αwe␈αhave␈αemployed␈αmost␈αheavily␈α
is
␈↓ α←␈↓the␈αability␈αto␈αmake␈αmultiple␈αuses␈αof␈αthe␈αknowledge␈αin␈αa␈αsingle␈αrepresentation.
␈↓ α←␈↓Rules, for example, are:

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?viewed as code and executed to drive the consultation,
␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?viewed␈α∪as␈α∪data␈α∪structures,␈α∪and␈α∪dissected␈α∪and␈α∪abstracted␈α∩to
␈↓ α←␈↓␈↓ β?form the rule models,
␈↓"β␈↓ α←␈↓␈↓ ββ(c)␈↓ β?dissected and examined to produce explanations,
␈↓"β␈↓ α←␈↓␈↓ ββ(d)␈↓ β?constructed during rule acquisition, and
␈↓"β␈↓ α←␈↓␈↓ ββ(e)␈↓ β?examined and reasoned about by the meta-rules.

␈↓ α←␈↓As␈α⊗another␈α↔example,␈α⊗the␈α↔schemata␈α⊗provide␈α⊗a␈α↔basis␈α⊗for␈α↔accessing␈α⊗and
␈↓ α←␈↓``understanding''␈α∂many␈α∂of␈α∂the␈α∂other␈α∂representations␈α∂used␈α∂in␈α∂the␈α∂system␈α∞and
␈↓ α←␈↓support␈αacquisition,␈αas␈αwell␈αas␈αstorage␈αand␈αretrieval.␈α Each␈αof␈α
these␈αalternate
␈↓ α←␈↓uses␈αrequires␈αknowledge␈αin␈αaddition␈α
to␈αthat␈αfound␈αin␈αthe␈α
representation,␈αbut
␈↓ α←␈↓if␈α
the␈α
knowledge␈αstored␈α
there␈α
can␈αbe␈α
decoded,␈α
it␈αneed␈α
not␈α
be␈α
represented␈αin
␈↓ α←␈↓more than one form.
␈↓"β␈↓ α←␈↓␈↓ β?It␈αis␈αimportant␈αto␈αnote␈αhere␈αthat␈αthe␈αfeasibility␈αof␈αsuch␈αmultiplicity␈αof
␈↓ α←␈↓uses␈α⊃is␈α⊃based␈α⊃less␈α⊃on␈α⊃the␈α⊃notion␈α⊂of␈α⊃production␈α⊃rules␈α⊃per␈α⊃se,␈α⊃than␈α⊃on␈α⊂the
␈↓ α←␈↓availability␈α
of␈α
a␈α
representation␈α
with␈α∞a␈α
␈↓↓small␈α
grain␈α
size␈↓␈α
and␈α
a␈α∞␈↓↓simple␈α
syntax
␈↓ α←␈↓↓and␈αsemantics␈↓.␈α ``Small'',␈αmodular␈αchunks␈αof␈αcode␈αwritten␈αin␈αa␈αsimple,␈αheavily
␈↓ α←␈↓stylized␈α⊂form␈α⊂(though␈α⊂not␈α⊃necessarily␈α⊂a␈α⊂situation-action␈α⊂form)␈α⊃would␈α⊂have
␈↓ α←␈↓done␈α∀as␈α∀well,␈α∀as␈α∀would␈α∪any␈α∀representation␈α∀with␈α∀simple␈α∀enough␈α∪internal
␈↓ α←␈↓structure␈αand␈αof␈αmanageable␈αsize.␈α The␈αintroduction␈αof␈αgreater␈αcomplexity␈αin
␈↓ α←␈↓the␈α∞representation,␈α∞or␈α∞the␈α∞use␈α
of␈α∞a␈α∞representation␈α∞that␈α∞encoded␈α
significantly
␈↓ α←␈↓larger␈α∂``chunks''␈α∞of␈α∂knowledge,␈α∂would␈α∞require␈α∂more␈α∂sophisticated␈α∞techniques
␈↓ α←␈↓for␈α
dissecting␈α
and␈α
manipulating␈αrepresentations␈α
than␈α
we␈α
have␈αdeveloped␈α
thus
␈↓ α←␈↓far.␈α But␈αthe␈αkey␈αlimitations␈αare␈αsize␈αand␈αcomplexity␈αof␈αstructure,␈αrather␈αthan
␈↓ α←␈↓a specific style of knowledge encoding.
␈↓"β␈↓ α←␈↓␈↓ β?It␈α∃is␈α∃interesting␈α∃to␈α∃consider␈α∃this␈α∃view␈α∃in␈α∃the␈α∃light␈α⊗of␈α∃historical
␈↓ α←␈↓developments.␈α≡ When␈α≡computers␈α≡were␈α≡programmed␈α≡with␈α≥plugboards,
␈↓ α←␈↓program␈α∪and␈α∩data␈α∪were␈α∪clearly␈α∩distinct.␈α∪ With␈α∪the␈α∩advent␈α∪of␈α∪the␈α∩stored
␈↓ α←␈↓program␈α⊂computer␈α∂to␈α⊂make␈α∂it␈α⊂possible,␈α∂and␈α⊂␈↓¬LISP␈↓␈α∂to␈α⊂make␈α∂it␈α⊂easy,␈α⊂the␈α∂idea
␈↓ α←␈↓arose␈α∀that␈α∪␈↓↓programs␈α∀could␈α∪be␈α∀considered␈α∪data␈↓.␈α∀ In␈α∪these␈α∀terms,␈α∀the␈α∪view
␈↓ α←␈↓proposed here adds to this the ideas that

␈↓"β␈↓ α←␈↓␈↓ ββ(a)␈↓ β?data␈αstructure␈αarchitecture␈α(i.e.,␈α
the␈αschemata)␈αcan␈αalso␈αbe␈α
data,
␈↓ α←␈↓␈↓ β?and
␈↓ α←␈↓␈↓268    CONCLUSIONS␈↓ 
#8-4␈↓

␈↓"β␈↓ α←␈↓␈↓ ββ(b)␈↓ β?control␈α∩architecture␈α∩(invocation␈α∩criteria,␈α∩the␈α∩meta-rules)␈α⊃can
␈↓ α←␈↓␈↓ β?also be data.

␈↓ α←␈↓As␈α⊗a␈α↔result,␈α⊗the␈α⊗system␈α↔gains␈α⊗access␈α⊗to␈α↔two␈α⊗more␈α⊗of␈α↔its␈α⊗fundamental
␈↓ α←␈↓components.
␈↓"β␈↓ α←␈↓␈↓ β?This␈α≤perspective␈α≤underlies␈α≤much␈α≤of␈α≤our␈α≤work␈α≥in␈α≤knowledge
␈↓ α←␈↓acquisition␈αand␈αknowledge␈αbase␈αmanagement␈α
and␈αseems␈αto␈αhave␈αpotential␈α
for
␈↓ α←␈↓future␈α
application.␈α
 For␈α
example,␈α
the␈αschemata␈α
can␈α
be␈α
viewed␈α
as␈αdefinitions
␈↓ α←␈↓of␈α∃data␈α∃structures,␈α∃and␈α∃that␈α⊗set␈α∃of␈α∃definitions␈α∃might␈α∃be␈α⊗examined␈α∃for
␈↓ α←␈↓inconsistencies.␈α∪ Do␈α∪two␈α∩different␈α∪structures␈α∪make␈α∪conflicting␈α∩assumptions
␈↓ α←␈↓about␈αthe␈αorganization␈αof␈αa␈αthird␈αstructure?␈α Are␈αthere␈αloops,␈αor␈αother␈αillegal
␈↓ α←␈↓topological␈αforms␈α
in␈αthe␈αschema␈α
hierarchy?␈α This␈α
problem␈αis␈αdifficult␈α
because
␈↓ α←␈↓it␈α~means,␈α~first,␈α→defining␈α~what␈α~``inconsistent''␈α→means␈α~for␈α~our␈α→informal
␈↓ α←␈↓knowledge␈α≠representation␈α≤and␈α≠then␈α≤devising␈α≠effective␈α≤procedures␈α≠for
␈↓ α←␈↓discovering it.
␈↓"β␈↓ α←␈↓␈↓ β?More␈α
speculatively,␈αmight␈α
the␈α
system␈αbe␈α
able␈αto␈α
discover␈α
or␈αconclude
␈↓ α←␈↓any␈α
``interesting␈α
things''␈α
in␈α
examining␈α
its␈α
own␈α
representations?␈α Two␈α
primitive
␈↓ α←␈↓forms␈αof␈αthis␈αexist␈αcurrently␈αin␈α␈↓¬TEIRESIAS␈↓: ␈αThe␈αrule␈αmodels␈αindicate␈αstatistical
␈↓ α←␈↓regularities␈α
found␈α
in␈α
the␈α
knowledge␈α∞base,␈α
and␈α
the␈α
schema␈α
network␈α∞editor␈α
is
␈↓ α←␈↓able␈α∩to␈α∩suggest␈α⊃possible␈α∩interconnections␈α∩for␈α∩a␈α⊃new␈α∩schema,␈α∩based␈α∩on␈α⊃its
␈↓ α←␈↓examination␈α∀of␈α∀the␈α∪existing␈α∀network.␈α∀ Can␈α∪more␈α∀sophisticated␈α∀forms␈α∪be
␈↓ α←␈↓developed?␈α~ Can␈α≠␈↓¬TEIRESIAS␈↓␈α~conceivably␈α≠make␈α~useful␈α≠suggestions␈α~about
␈↓ α←␈↓representation␈α
design,␈α∞by␈α
noticing,␈α
perhaps,␈α∞that␈α
(a)  two␈α∞representations␈α
are
␈↓ α←␈↓very␈α→``similar'',␈α→(b)  there␈α_are␈α→relatively␈α→``few''␈α_instances␈α→of␈α→either,␈α_and
␈↓ α←␈↓(c)  combining␈α
the␈α
two␈α
might␈α
make␈α
the␈α
schema␈α
network␈α
``simpler.'' ␈αWhile␈α
this
␈↓ α←␈↓would␈α∂have␈α∂great␈α∂utility␈α∂in␈α∞dealing␈α∂with␈α∂the␈α∂organization␈α∂and␈α∂structure␈α∞of
␈↓ α←␈↓large␈α
programs,␈α
the␈α
difficulty␈α
starts␈αwith␈α
attempts␈α
to␈α
give␈α
precise␈αdefinitions
␈↓ α←␈↓to␈α∞the␈α∂words␈α∞in␈α∞quotes,␈α∂and␈α∞multiplies␈α∂quickly.␈α∞ But␈α∞let␈α∂us␈α∞speculate␈α∂just␈α∞a
␈↓ α←␈↓little␈α⊃further.␈α⊃ Is␈α⊂it␈α⊃possible␈α⊃that␈α⊃some␈α⊂of␈α⊃those␈α⊃suggestions␈α⊃from␈α⊂␈↓¬TEIRESIAS␈↓
␈↓ α←␈↓about␈α
data␈αstructures␈α
might␈α
in␈αfact␈α
have␈αimplications␈α
for␈α
understanding␈αthe
␈↓ α←␈↓domain␈α∀itself? ␈α∀Can␈α∃the␈α∀regularities␈α∀discovered␈α∃by␈α∀the␈α∀rule␈α∃models,␈α∀for
␈↓ α←␈↓instance,␈α∪ever␈α∪suggest␈α∪new␈α∪and␈α∪interesting␈α∪things␈α∪about␈α∪the␈α∪structure␈α∩of
␈↓ α←␈↓reasoning␈α⊃in␈α∩the␈α⊃domain? ␈α⊃Or␈α∩can␈α⊃the␈α⊃system's␈α∩ability␈α⊃to␈α∩discover␈α⊃useful
␈↓ α←␈↓things␈α~about␈α≠data␈α~structures␈α≠be␈α~made␈α≠sophisticated␈α~enough␈α≠that␈α~its
␈↓ α←␈↓suggestions offer some help in forming theories about the domain?
␈↓"β␈↓ α←␈↓␈↓ β?These␈α∪are␈α∀difficult␈α∪issues.␈α∀ They␈α∪address␈α∀the␈α∪problem␈α∀of␈α∪pattern
␈↓ α←␈↓detection␈α∂and␈α∂theory␈α∂formation␈α∞applied␈α∂to␈α∂a␈α∂slightly␈α∂different␈α∞domain--the
␈↓ α←␈↓domain␈α∀of␈α∀representations.␈α∀ Chapter␈α∪5␈α∀considered␈α∀this␈α∀issue␈α∀briefly␈α∪and
␈↓ α←␈↓examined␈α∞the␈α∞rule␈α∞models␈α∞as␈α∂an␈α∞example␈α∞of␈α∞concept␈α∞formation.␈α∞ All␈α∂of␈α∞the
␈↓ α←␈↓hard␈α∂problems␈α⊂noted␈α∂there␈α∂are␈α⊂yet␈α∂to␈α⊂be␈α∂solved,␈α∂as␈α⊂are␈α∂the␈α⊂more␈α∂complex
␈↓ α←␈↓issues␈α⊃involved␈α⊂in␈α⊃theory␈α⊃formation.␈α⊂ The␈α⊃examples␈α⊂found␈α⊃in␈α⊃the␈α⊂present
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓ system serve basically as demonstrations of feasibility.

␈↓ α←␈↓␈↓αTheme 4:  Programs can be self-understanding.␈↓
␈↓"β␈↓ α←␈↓␈↓ β?One␈αaspect␈αof␈αthe␈αidea␈αof␈α``self-understanding''␈αhas␈αbeen␈αillustrated␈αin
␈↓ α←␈↓␈↓8-4␈↓ ¬yTHE OTHER THEMES; SOME SPECULATIONS    269␈↓

␈↓"β␈↓ α←␈↓the␈αprevious␈αsection␈αand␈αconcerns␈α
a␈αprogram's␈αability␈αto␈αunderstand␈α
itself␈αin
␈↓ α←␈↓terms␈αof␈αrepresentations␈αand␈αdata␈αstructures.␈α Here␈αwe␈αconsider␈αhow␈αit␈αmight
␈↓ α←␈↓understand its own behavior.
␈↓"β␈↓ α←␈↓␈↓ β?To␈αmake␈αthe␈αdiscussion␈αmore␈αconcrete,␈αconsider␈αthe␈αfollowing␈αcriteria
␈↓ α←␈↓for ``understanding'':

␈↓"β␈↓ α←␈↓␈↓ ββCan the program
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?explain its behavior?
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?explain what it knows?
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?change its behavior?
␈↓"β␈↓ α←␈↓␈↓ ββ␈↓ β?change the content of its knowledge?

␈↓"β␈↓ α←␈↓␈↓ β?These␈α
criteria␈α
can␈αbe␈α
met␈α
with␈α
varying␈αlevels␈α
of␈α
success␈α
with␈αrespect
␈↓ α←␈↓to␈α
completeness␈α
(i.e.,␈α
how␈α
much␈α
of␈α
the␈α
system␈α
can␈α
be␈α
explained␈α
or␈α
changed)
␈↓ α←␈↓and␈αgenerality␈α(i.e.,␈αare␈αthe␈αtechniques␈αspecific␈αto␈αthe␈αcurrent␈αimplementation
␈↓ α←␈↓or are they more broadly applicable).
␈↓"β␈↓ α←␈↓␈↓ β?Consider␈α∩the␈α∪issue␈α∩of␈α∩generality␈α∪first,␈α∩and␈α∩examine␈α∪the␈α∩combined
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓-␈↓¬MYCIN␈↓␈α∀system␈α∀as␈α∃one␈α∀example.␈α∀ Chapter␈α∃3␈α∀noted␈α∀the␈α∃variety␈α∀of
␈↓ α←␈↓techniques␈αthat␈αhas␈αbeen␈αused␈αto␈αgenerate␈αexplanations.␈α These␈αranged␈αfrom
␈↓ α←␈↓one␈α∞program␈α∞written␈α∂specifically␈α∞to␈α∞explain␈α∞part␈α∂of␈α∞another,␈α∞to␈α∂the␈α∞system's
␈↓ α←␈↓ability␈αto␈αuse␈αthe␈αtemplates␈αto␈αexamine␈αthe␈αsame␈αrules␈αit␈αwas␈αexecuting.␈α This
␈↓ α←␈↓latter␈α∂form␈α∂of␈α∂introspection␈α∂is␈α∂the␈α∂most␈α∂powerful␈α∂of␈α∂the␈α∂solutions␈α⊂we␈α∂have
␈↓ α←␈↓explored,␈α~since␈α~it␈α→adjusts␈α~automatically␈α~to␈α→changes␈α~in␈α~the␈α→program's
␈↓ α←␈↓knowledge␈α
base␈α
and␈α
control␈α
structure␈α
(at␈α
least␈α
the␈α
part␈α
of␈α
the␈α
control␈α
structure
␈↓ α←␈↓that can be captured in meta-rules).
␈↓"β␈↓ α←␈↓␈↓ β?The␈αquestion␈αof␈αcompleteness␈αfor␈αthe␈αfirst␈αthree␈αof␈αthe␈αcriteria␈αabove
␈↓ α←␈↓is␈α
answered␈α
currently␈αby␈α
considering␈α
how␈α
much␈αknowledge␈α
is␈α
in␈α
rules.␈α The
␈↓ α←␈↓system␈α∞can␈α∞explain␈α∞behavior␈α∞that␈α∂results␈α∞from␈α∞rule␈α∞invocation,␈α∞can␈α∂refer␈α∞to
␈↓ α←␈↓knowledge␈α
embedded␈α
in␈α
rules,␈α
and␈α
can␈α
change␈α
its␈α
behavior␈α
by␈α
the␈α
acquisition
␈↓ α←␈↓of␈α⊂new␈α⊃rules.␈α⊂ Where␈α⊂knowledge␈α⊃or␈α⊂behavior␈α⊂is␈α⊃not␈α⊂rule␈α⊂based,␈α⊃it␈α⊂cannot
␈↓ α←␈↓currently␈α
be␈α
explained.␈α
 (The␈αability␈α
to␈α
acquire␈α
new␈α
conceptual␈αprimitives␈α
via
␈↓ α←␈↓the␈αschemata␈αoffers␈αanother␈αfoundation␈αfor␈αacquisition␈αof␈αnew␈αknowledge,␈αso
␈↓ α←␈↓the␈αsystem␈αis␈αcurrently␈αmore␈αcomplete␈αwith␈αregard␈αto␈αchanging␈αthe␈αcontent␈αof
␈↓ α←␈↓its knowledge.)
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∂appears␈α∂to␈α∂be␈α∂a␈α∂fundamental␈α∂limit␈α∂on␈α∂the␈α∂completeness␈α∂with
␈↓ α←␈↓which␈α⊂program␈α⊂behavior␈α⊂and␈α⊃knowledge␈α⊂might␈α⊂plausibly␈α⊂be␈α⊃explained␈α⊂or
␈↓ α←␈↓modified.␈α∃ It␈α∃is␈α∀characterized␈α∃in␈α∃the␈α∀introduction␈α∃to␈α∃[Minksy68]␈α∃as␈α∀the
␈↓ α←␈↓distinction␈α↔between␈α↔``compiled''␈α↔and␈α↔``interpreted''␈α↔behavior,␈α↔where␈α_it␈α↔is
␈↓ α←␈↓suggested␈α
that␈α
some␈α
behavior,␈α
some␈α
knowledge,␈α
may␈α
reasonably␈αbe␈α
considered
␈↓ α←␈↓``compiled''␈α⊂and,␈α∂hence,␈α⊂inaccessible␈α∂to␈α⊂further␈α∂decomposition.␈α⊂ The␈α⊂issue␈α∂is
␈↓ α←␈↓more␈α
than␈α
just␈α
choosing␈α
a␈αconvenient␈α
level␈α
at␈α
which␈α
to␈α
view␈αsystem␈α
behavior,
␈↓ α←␈↓as␈αwe␈αdid␈αin␈αchapter␈α3.␈α Some␈αbehavior␈αand␈αknowledge␈αis␈αsimply␈αgoing␈αto␈αbe
␈↓ α←␈↓inexplicable.␈α∪ For␈α∪␈↓¬TEIRESIAS␈↓-␈↓¬MYCIN␈↓,␈α∪for␈α∪example,␈α∪this␈α∪might␈α∪mean␈α∪that␈α∪the
␈↓ α←␈↓system␈α∞would␈α
not␈α∞and,␈α
indeed,␈α∞need␈α
not␈α∞ever␈α
be␈α∞able␈α
to␈α∞explain␈α∞things␈α
like
␈↓ α←␈↓the mechanism of rule retrieval.
␈↓ α←␈↓␈↓270    CONCLUSIONS␈↓ 
#8-4␈↓

␈↓"β␈↓ α←␈↓␈↓ β?In␈α∂line␈α∞with␈α∂our␈α∂emphasis␈α∞on␈α∂making␈α∞knowledge␈α∂explicit,␈α∂we␈α∞might
␈↓ α←␈↓add␈α⊗another␈α⊗possibility: ␈α∃In␈α⊗addition␈α⊗to␈α∃being␈α⊗compiled␈α⊗or␈α∃interpreted,
␈↓ α←␈↓knowledge␈α∞might␈α∞also␈α∞simply␈α∞be␈α∞absent.␈α∞ A␈α∞program␈α∞clearly␈α∂cannot␈α∞explain
␈↓ α←␈↓what␈αit␈αdoesn't␈αknow,␈αand␈αbefore␈αthe␈αuse␈αof␈αmeta-rules,␈α␈↓¬MYCIN␈↓␈αhad␈αno␈αway␈αof
␈↓ α←␈↓explaining␈αwhy␈αcertain␈αrules␈α
were␈αexecuted␈αin␈αa␈α
particular␈αorder.␈α It␈αwas␈αin␈α
a
␈↓ α←␈↓position␈α∀roughly␈α∀analogous␈α∀to␈α∀someone␈α∪who␈α∀can␈α∀recall␈α∀a␈α∀list␈α∀from␈α∪rote
␈↓ α←␈↓memorization, but never learned the rationale behind its order.␈↓
5␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈αnoted␈αearlier,␈αour␈αefforts␈αtoward␈αexplanation,␈αand␈αthis␈α
larger␈αgoal
␈↓ α←␈↓of␈α⊂self-understanding,␈α⊂have␈α⊂really␈α⊂only␈α∂begun.␈α⊂ The␈α⊂issue␈α⊂is␈α⊂again␈α⊂one␈α∂of
␈↓ α←␈↓representation:␈α
How␈α
can␈α
knowledge␈α
be␈α
represented␈α
in␈α
forms␈α
that␈α
allow␈α∞it␈α
to
␈↓ α←␈↓be␈α
both␈αused␈α
and␈αdissected? ␈α
Our␈αuse␈α
of␈αrules␈α
as␈αa␈α
high-level␈α
language␈αhas
␈↓ α←␈↓provided␈αa␈αcertain␈αlevel␈αof␈αperformance,␈αbut␈αas␈αprevious␈αsections␈αhave␈α
noted,
␈↓ α←␈↓there are many problems yet to be solved.

␈↓ α←␈↓␈↓αTheme␈α
5: ␈α
A␈αrepresentation␈α
can␈α
usefully␈α
be␈αmore␈α
than␈α
a␈α
densely␈αencoded
␈↓ α←␈↓αstring of bits.␈↓
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∂is␈α⊂an␈α∂obvious␈α∂corollary␈α⊂to␈α∂the␈α∂belief␈α⊂that␈α∂a␈α∂program␈α⊂can␈α∂be
␈↓ α←␈↓more␈αthan␈αthe␈αminimum␈αset␈αof␈αsymbol␈αmanipulation␈αoperations␈αnecessary␈αfor
␈↓ α←␈↓solving␈α⊂a␈α⊂problem.␈α⊂ Analogously,␈α⊂a␈α∂representation␈α⊂can␈α⊂be␈α⊂viewed␈α⊂in␈α∂terms
␈↓ α←␈↓other␈α⊃than␈α∩the␈α⊃minimum␈α⊃number␈α∩of␈α⊃bits␈α⊃that␈α∩will␈α⊃carry␈α∩the␈α⊃information
␈↓ α←␈↓required.␈α∪ This␈α∀becomes␈α∪necessary␈α∀if␈α∪we␈α∀want␈α∪a␈α∀program␈α∪to␈α∀modify␈α∪its
␈↓ α←␈↓knowledge␈αbase␈αwithout␈αintroducing␈αerrors: ␈αIf␈αthe␈αprogram␈αis␈αto␈αmanipulate
␈↓ α←␈↓its␈α
representations␈α
in␈α
intelligent␈αways,␈α
then␈α
it␈α
needs␈αto␈α
know␈α
a␈α
good␈αdeal␈α
more
␈↓ α←␈↓about each than standard declarations typically supply.
␈↓"β␈↓ α←␈↓␈↓ β?Some␈α⊃of␈α⊃the␈α⊃additional␈α⊃information␈α⊃about␈α⊃representations␈α⊃we␈α⊂used
␈↓ α←␈↓includes: ␈α⊃its␈α⊃structure,␈α⊃the␈α⊂set␈α⊃of␈α⊃legal␈α⊃values␈α⊂for␈α⊃each␈α⊃component␈α⊃of␈α⊂the
␈↓ α←␈↓structure,␈α⊃its␈α⊃interrelations␈α⊃with␈α⊂other␈α⊃representations␈α⊃(as␈α⊃indicated␈α⊃by␈α⊂the
␈↓ α←␈↓␈↓∧RELATIONS␈↓),␈α∪and␈α∪its␈α∪place␈α∪in␈α∩the␈α∪whole␈α∪collection␈α∪of␈α∪data␈α∪structures␈α∩(as
␈↓ α←␈↓indicated␈α∞by␈α
the␈α∞schema␈α∞hierarchy).␈α
 All␈α∞of␈α∞this␈α
information␈α∞is␈α∞necessary␈α
to
␈↓ α←␈↓perform␈α↔even␈α⊗the␈α↔simple␈α⊗and␈α↔straightforward␈α⊗``clerical''␈α↔kinds␈α↔of␈α⊗tasks
␈↓ α←␈↓involved in knowledge base maintenance.
␈↓"β␈↓ α←␈↓␈↓ β?In␈αaddition,␈αit␈αmakes␈αpossible␈αmuch␈αof␈αthe␈α``intelligence''␈αdisplayed␈αby
␈↓ α←␈↓the␈α∩acquisition␈α∪system.␈α∩ As␈α∪noted␈α∩above,␈α∩the␈α∪representations␈α∩used␈α∪in␈α∩the
␈↓ α←␈↓knowledge␈α⊃base␈α∩are␈α⊃both␈α∩data␈α⊃structures␈α⊃and␈α∩a␈α⊃model␈α∩of␈α⊃the␈α∩domain␈α⊃of
␈↓ α←␈↓application.␈α⊃ Part␈α⊃of␈α⊃an␈α⊂``intelligent''␈α⊃process␈α⊃of␈α⊃acquisition␈α⊃involves␈α⊂being
␈↓ α←␈↓able␈α
both␈α
to␈α
use␈α
this␈α
information␈α
and␈α
to␈α
insure␈α
that␈α
it␈α
is␈αproperly␈α
maintained.
␈↓ α←␈↓Chapter␈α5,␈αfor␈αexample,␈αnoted␈αthat␈αone␈αof␈αthe␈αconsistency␈αconstraints␈αused␈αin
␈↓ α←␈↓generating␈αinterpretations␈αof␈αa␈αnew␈αrule␈αrelied␈αon␈αthis␈αworld␈αmodel␈αto␈αinsure
␈↓ α←␈↓that␈α⊂an␈α⊂attribute␈α∂in␈α⊂a␈α⊂clause␈α∂was␈α⊂paired␈α⊂with␈α∂an␈α⊂appropriate␈α⊂value.␈α∂ The
␈↓ α←␈↓integrity␈α∩of␈α∩the␈α∩model␈α∩is␈α∩maintained␈α∩by␈α∩devices␈α∩like␈α∩the␈α∩␈↓∧RELATIONS␈↓␈α⊃list,
␈↓ α←␈↓which␈α∞insures␈α∞that␈α∞new␈α∞structures␈α
are␈α∞properly␈α∞integrated␈α∞and␈α∞that␈α
existing
␈↓ α←␈↓dependencies are maintained.

␈↓"β␈↓ α←␈↓_______________________________
␈↓"β␈↓ α←␈↓[5]␈α
See␈α[Parnas75]␈α
for␈αsome␈α
further␈α
thoughts␈αon␈α
``compiled''␈αand␈α
``interpreted''
␈↓ α←␈↓knowledge and its use in system design.
␈↓ α←␈↓␈↓8-4␈↓ ¬yTHE OTHER THEMES; SOME SPECULATIONS    271␈↓

␈↓"β␈↓ α←␈↓␈↓ β?All␈α∩of␈α∩this␈α∩means␈α∩an␈α⊃emphasis␈α∩on␈α∩the␈α∩careful␈α∩representation␈α⊃and
␈↓ α←␈↓organization␈α⊂of␈α⊂this␈α⊂information.␈α⊂ The␈α∂schema␈α⊂language␈α⊂is␈α⊂one␈α⊂attempt␈α∂to
␈↓ α←␈↓provide␈αa␈αframework␈αfor␈αorganizing␈αand␈αaccessing␈αthis␈αinformation␈αand␈αmay
␈↓ α←␈↓help to make its expression easier.

␈↓"β␈↓ α←␈↓␈↓ β?These␈α
first␈α
five␈α∞themes␈α
have␈α
explored␈α∞how␈α
␈↓¬TEIRESIAS␈↓␈α
supplies␈α∞a␈α
base
␈↓ α←␈↓of␈αmeta-level␈αknowledge--information␈αabout␈αthe␈αorganization,␈α
structure,␈αand
␈↓ α←␈↓use␈αof␈αknowledge␈αin␈αthe␈αperformance␈α
program.␈α The␈αultimate␈αgoal␈αis␈αto␈α
make
␈↓ α←␈↓possible␈α⊃a␈α⊃useful␈α⊃level␈α⊃of␈α⊃introspection,␈α⊃one␈α⊃that␈α⊃will␈α⊃support␈α⊃``intelligent''
␈↓ α←␈↓behavior␈α∂for␈α∂the␈α∂acquisition,␈α∂explanation,␈α∂and␈α∂effective␈α∂use␈α⊂of␈α∂object-level
␈↓ α←␈↓knowledge.␈α∂ ␈↓↓Task-specific␈α∂high-level␈α∞languages␈α∂␈↓are␈α∂the␈α∂primary␈α∞mechanism
␈↓ α←␈↓used␈α
to␈α
make␈α
this␈α
feasible,␈α
since␈α
they␈α
avoid␈α
the␈α
difficult␈α
problems␈α
involved␈α
in
␈↓ α←␈↓having␈α
a␈α∞program␈α
read␈α∞ordinary␈α
code␈α
or␈α∞examine␈α
standard␈α∞data␈α
structures.
␈↓ α←␈↓The␈α∀emphasis␈α∀on␈α∀␈↓↓making␈α∪knowledge␈α∀in␈α∀programs␈α∀explicit␈α∀and␈α∪accessible␈↓
␈↓ α←␈↓follows␈αfrom␈αthe␈αdesire␈αto␈αhave␈αthe␈αprogram␈αmodify␈αits␈αown␈αknowledge␈αbase
␈↓ α←␈↓and␈α∞behavior.␈α∂ To␈α∞avoid␈α∂introducing␈α∞errors␈α∞in␈α∂doing␈α∞this,␈α∂it␈α∞needs␈α∂a␈α∞large
␈↓ α←␈↓store␈α
of␈α␈↓↓information␈α
about␈α
its␈αown␈α
representations␈↓,␈α
information␈αof␈α
the␈αsort␈α
that
␈↓ α←␈↓is␈α
typically␈α
left␈αeither␈α
informally␈α
specified␈α
or␈αomitted␈α
entirely.␈α
 The␈αconcept␈α
of
␈↓ α←␈↓a␈α∞␈↓↓program␈α∞that␈α∞``understands''␈α∞itself␈↓␈α∞concerns␈α∞its␈α∞ability␈α∞to␈α∞account␈α∞for,␈α∞and
␈↓ α←␈↓modify,␈α∞its␈α∞behavior␈α∞and␈α∂knowledge;␈α∞this␈α∞is␈α∞introspection␈α∞at␈α∂the␈α∞behavioral
␈↓ α←␈↓level.␈α Finally,␈αthe␈α
suggestion␈αwas␈αmade␈α
that␈αa␈α␈↓↓representation␈α
can␈αbe␈αviewed␈α
in
␈↓ α←␈↓↓a␈α
broad␈α
context␈↓,␈α
both␈α
as␈α
a␈α
data␈α
structure␈α
in␈α
the␈α
program␈α
and␈α
as␈α
part␈α
of␈α
the
␈↓ α←␈↓program's␈α∩model␈α∩of␈α∩the␈α∩world.␈α∩ This␈α∪context␈α∩is␈α∩used␈α∩as␈α∩the␈α∩basis␈α∪for␈α∩a
␈↓ α←␈↓language␈α⊃and␈α⊃framework␈α⊃in␈α⊃which␈α⊃to␈α⊃organize␈α⊃and␈α⊃express␈α⊃much␈α∩of␈α⊃the
␈↓ α←␈↓required knowledge.

␈↓ α←␈↓␈↓αTheme 6:  A program can have some grasp on its own complexity.␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αthought␈αhas␈αbeen␈αaround␈αfor␈α
some␈αtime␈αthat␈αa␈αprogram␈αought␈α
to
␈↓ α←␈↓be␈α
able␈α
to␈αhelp␈α
out␈α
in␈αdealing␈α
with␈α
its␈α
own␈αcomplexities␈α
(a␈α
recent␈αcollection␈α
of
␈↓ α←␈↓suggestions␈αalong␈αthis␈αline␈αis␈αin␈α[Winograd74]).␈α There␈αare␈αat␈αleast␈αtwo␈αways
␈↓ α←␈↓in which it should be possible to enlist a program's help.
␈↓"β␈↓ α←␈↓␈↓ β?The␈α
first␈α
involves␈α
simply␈α
coping␈α
with␈α
the␈α
complexity␈α
that␈α
inevitably
␈↓ α←␈↓arises.␈α It␈αis,␈αas␈α
we␈αhave␈αseen,␈αdifficult␈αto␈α
make␈αa␈αsingle␈αsmall␈αchange␈α
to␈αany
␈↓ α←␈↓large␈α
system.␈αThere␈α
are␈α
typically␈αa␈α
number␈αof␈α
related␈α
changes␈αthat␈α
have␈αto␈α
be
␈↓ α←␈↓made␈αand␈αconstraints␈αthat␈αmust␈αbe␈αmet.␈α Rather␈αthan␈αrelying␈αon␈αmemory␈α(or
␈↓ α←␈↓documentation)␈α
to␈α
keep␈α
track␈αof␈α
all␈α
the␈α
details,␈α
why␈αnot␈α
have␈α
the␈α
system␈αdo␈α
it.
␈↓ α←␈↓Whatever␈α∩knowledge␈α⊃there␈α∩is␈α∩concerning␈α⊃those␈α∩details␈α⊃ought␈α∩to␈α∩be␈α⊃made
␈↓ α←␈↓explicit␈α∪and␈α∪then␈α∩made␈α∪accessible␈α∪to␈α∩and␈α∪comprehensible␈α∪by␈α∪the␈α∩system.
␈↓ α←␈↓Several␈αexamples␈αof␈αthis␈αhave␈αbeen␈αexplored␈αin␈αearlier␈αchapters.␈α The␈αuse␈αof
␈↓ α←␈↓the␈α∪schemata␈α∪for␈α∀acquisition␈α∪of␈α∪new␈α∀conceptual␈α∪primitives␈α∪is␈α∀a␈α∪primary
␈↓ α←␈↓example.␈α∞ The␈α∂schemata␈α∞were␈α∂designed␈α∞to␈α∞confront␈α∂exactly␈α∞this␈α∂problem␈α∞of
␈↓ α←␈↓the␈α_complexity␈α_and␈α_detail␈α_of␈α_structure␈α_and␈α→interrelationships␈α_typically
␈↓ α←␈↓encountered␈αin␈α
any␈αlarge␈α
set␈αof␈α
data␈αstructures.␈α
 Another␈αexample␈α
is␈αfound␈α
in
␈↓ α←␈↓the␈α∂suggestion␈α∂that␈α∂it␈α∂ought␈α∂to␈α⊂be␈α∂possible␈α∂to␈α∂edit␈α∂the␈α∂schemata␈α⊂and␈α∂have
␈↓ α←␈↓␈↓¬TEIRESIAS␈↓␈α
edit␈α
all␈α
the␈α
instances␈α
similarly.␈α
 The␈α
use␈α
of␈α
a␈α
``totally␈α
typed''␈α
language
␈↓ α←␈↓␈↓272    CONCLUSIONS␈↓ 
#8-4␈↓

␈↓"β␈↓ α←␈↓during␈α∩acquisition␈α∩is␈α∩still␈α∩another␈α∩example.␈α∩ As␈α∩illustrated,␈α∩this␈α∪makes␈α∩it
␈↓ α←␈↓possible␈α
for␈α
the␈α
system␈α
to␈α
specify␈α
to␈α
the␈α
expert␈α
the␈α
legal␈α
form␈α
of␈α
the␈αanswer␈α
to
␈↓ α←␈↓any␈αof␈αits␈αquestions␈α
and␈αto␈αcheck␈αthe␈α
responses␈αfor␈αsyntactic␈αvalidity.␈α
 Finally,
␈↓ α←␈↓even␈α
the␈α
simple␈α
recap␈α
provided␈α
by␈α
the␈α
explanation␈α
routines␈α
can␈α
help␈αthe␈α
user
␈↓ α←␈↓cope with the wealth of detail that arises in a long consultation.
␈↓"β␈↓ α←␈↓␈↓ β?The␈αsecond␈αform␈αof␈αaid␈αmight␈αarise␈αfrom␈αthe␈αcapability␈αof␈αthe␈αsystem
␈↓ α←␈↓to␈αsimplify␈αitself␈αfor␈αthe␈αbenefit␈αof␈αthe␈αviewer.␈α Introspection␈αmay␈αmean␈αlittle
␈↓ α←␈↓if␈α
a␈αprogram␈α
is␈α
very␈αlarge␈α
and␈αits␈α
behavior␈α
complex.␈α If␈α
the␈αsystem␈α
``explains''
␈↓ α←␈↓itself␈α∞at␈α
a␈α∞level␈α
that␈α∞leaves␈α
the␈α∞viewer␈α
drowning␈α∞in␈α
a␈α∞wealth␈α
of␈α∞detail,␈α
little
␈↓ α←␈↓has␈α∞been␈α∞accomplished.␈α∞ Is␈α∞it␈α∞possible,␈α∞then,␈α∞for␈α∞a␈α∞program␈α∞to␈α∂abstract␈α∞that
␈↓ α←␈↓detail␈α⊃and␈α⊂present␈α⊃a␈α⊂comprehensible␈α⊃picture? ␈α⊂Can␈α⊃the␈α⊃``magnification''␈α⊂be
␈↓ α←␈↓varied to provide accounts which offer different levels of details?
␈↓"β␈↓ α←␈↓␈↓ β?Some␈α
very␈α∞basic␈α
steps␈α∞in␈α
this␈α∞direction␈α
were␈α∞described␈α
in␈α∞chapters␈α
3
␈↓ α←␈↓and␈α_5.␈α↔ The␈α_``higher␈α↔level␈α_explanations''␈α↔available␈α_with␈α↔the␈α_␈↓∧WHY␈α↔<n>␈↓
␈↓ α←␈↓command,␈α∀for␈α∃instance,␈α∀offer␈α∃a␈α∀very␈α∀simple␈α∃form␈α∀of␈α∃simplification␈α∀and
␈↓ α←␈↓variable␈α
magnification.␈α∞ The␈α
rule␈α
models␈α∞are␈α
another␈α
example: ␈α∞As␈α
abstract
␈↓ α←␈↓descriptions␈αof␈αthe␈α
content␈αof␈αthe␈α
knowledge␈αbase␈αthey␈α
help␈αto␈αmake␈αclear␈α
the
␈↓ α←␈↓general trends in the reasoning while suppressing less important detail.
␈↓"β␈↓ α←␈↓␈↓ β?But␈α∂these␈α∞are␈α∂only␈α∞first␈α∂steps.␈α∞ A␈α∂more␈α∞sophisticated␈α∂solution␈α∞would
␈↓ α←␈↓make␈α∂it␈α∂possible␈α⊂to␈α∂base␈α∂different␈α∂explanations␈α⊂on␈α∂different␈α∂models␈α⊂of␈α∂the
␈↓ α←␈↓underlying␈α∂processes.␈α∂ It␈α∂might␈α∂even␈α∂be␈α∞able␈α∂to␈α∂choose␈α∂a␈α∂model␈α∂tailored␈α∞to
␈↓ α←␈↓both␈αthe␈αcurrent␈αsequence␈αof␈αevents␈αand␈αthe␈αcurrent␈αviewer.␈α There␈αis␈αclearly
␈↓ α←␈↓much work that remains to be done.

␈↓ α←␈↓␈↓αTheme 7: Programs can be self-adjusting.␈↓
␈↓"β␈↓ α←␈↓␈↓ β?As␈α↔the␈α⊗previous␈α↔discussion␈α⊗suggests,␈α↔dealing␈α⊗with␈α↔complexity␈α⊗is
␈↓ α←␈↓difficult.␈α∞ One␈α
approach␈α∞to␈α∞the␈α
problem␈α∞views␈α
it␈α∞as␈α∞an␈α
issue␈α∞of␈α∞design␈α
and
␈↓ α←␈↓asks: ␈αRather␈α
than␈αcreating␈αtools␈α
to␈αrecapture␈α
programs␈αwhose␈αcomplexity␈α
has
␈↓ α←␈↓gotten␈α∂out␈α∂of␈α∂hand,␈α∂is␈α∂it␈α∂instead␈α∂possible␈α∂to␈α∂design␈α∂programs␈α∂in␈α∂ways␈α∂that
␈↓ α←␈↓avoid␈αcomplexity␈αfrom␈αthe␈αoutset? ␈αThat␈αis,␈αcan␈αwe␈αtry␈αto␈αpartition␈αproblems
␈↓ α←␈↓and␈α↔design␈α⊗representations␈α↔in␈α↔ways␈α⊗that␈α↔make␈α⊗them␈α↔simple␈α↔and␈α⊗self-
␈↓ α←␈↓contained.␈α
 A␈α
few␈αrelevant␈α
ideas␈α
have␈αbeen␈α
in␈α
the␈αair␈α
for␈α
some␈αtime.␈α
 Simon's
␈↓ α←␈↓␈↓↓Sciences␈α∞of␈α
the␈α∞Artificial␈↓␈α∞contains␈α
some␈α∞very␈α
general␈α∞observations␈α∞on␈α
system
␈↓ α←␈↓decomposition,␈α→while␈α→␈↓↓Structured␈α_Programming␈↓␈α→[Dijkstra72]␈α→offers␈α_some
␈↓ α←␈↓thoughts on control structure design.
␈↓"β␈↓ α←␈↓␈↓ β?There␈α∀are␈α∀also␈α∀some␈α∀well-established␈α∀techniques␈α∀of␈α∪programming
␈↓ α←␈↓which␈α
confront␈α
this␈α
problem.␈α
 One␈α
approach␈α
emphasizes␈α
dealing␈αwith␈α
change
␈↓ α←␈↓by␈α⊗␈↓↓insulating␈↓␈α⊗one␈α↔part␈α⊗of␈α⊗the␈α⊗program␈α↔from␈α⊗another.␈α⊗ A␈α↔number␈α⊗of
␈↓ α←␈↓techniques␈α∞have␈α∞been␈α
used,␈α∞and␈α∞a␈α∞somewhat␈α
larger␈α∞number␈α∞of␈α∞names␈α
have
␈↓ α←␈↓been invented for them.

␈↓"β␈↓ α←␈↓␈↓ ββ(1)␈↓ β?␈↓↓Indirection␈↓.  Indirection␈αinsulates␈αthe␈αprogram␈αfrom␈αreferential
␈↓ α←␈↓␈↓ β?changes.␈α
 Thus,␈α
indirection␈α
in␈α
code␈α
(indirect␈α
addressing)␈α
makes
␈↓ α←␈↓␈↓ β?possible␈α∞compile␈α∞time␈α∞(or␈α∞even␈α∞execution␈α∞time)␈α∂assignment␈α∞of
␈↓ α←␈↓␈↓ β?addresses.␈α∞ Indirection␈α∞in␈α∞data␈α∞(e.g.,␈α∂use␈α∞of␈α∞a␈α∞pointer␈α∞to␈α∂a␈α∞set
␈↓ α←␈↓␈↓8-4␈↓ ¬yTHE OTHER THEMES; SOME SPECULATIONS    273␈↓

␈↓"β␈↓ α←␈↓␈↓ β?rather␈α⊂than␈α⊂an␈α⊂explicit␈α⊂list␈α⊂of␈α⊂its␈α⊂members)␈α⊂offers␈α⊃a␈α⊂similar
␈↓ α←␈↓␈↓ β?flexibility.

␈↓"β␈↓ α←␈↓␈↓ ββ(2)␈↓ β?␈↓↓Records␈↓.  One␈α≠level␈α≠of␈α~record␈α≠structures␈α≠will␈α≠insulate␈α~a
␈↓ α←␈↓␈↓ β?program␈α→from␈α→changes␈α→in␈α→data␈α→structure␈α→implementation
␈↓ α←␈↓␈↓ β?(``dataless␈α![representation-free]␈α!programming''␈α [Balzer67]).
␈↓ α←␈↓␈↓ β?Two␈α∪levels␈α∪of␈α∩record␈α∪structures␈α∪can␈α∩be␈α∪used␈α∪to␈α∪insulate␈α∩a
␈↓ α←␈↓␈↓ β?program␈α∃from␈α∃certain␈α∃kinds␈α∀of␈α∃changes␈α∃in␈α∃data␈α∀structure
␈↓ α←␈↓␈↓ β?design (``data independence,'' [Bachman75]).

␈↓"β␈↓ α←␈↓␈↓ ββ(3)␈↓ β?␈↓↓Data␈αabstraction␈↓.  Where␈αrecords␈αdefine␈αa␈αrepresentation␈αby␈α
its
␈↓ α←␈↓␈↓ β?structure,␈αdata␈αabstraction␈αdefines␈αit␈αby␈αits␈αbehavior.␈α It␈αoffers
␈↓ α←␈↓␈↓ β?a␈α⊂similar␈α⊂form␈α⊂of␈α⊃insulation,␈α⊂allowing␈α⊂code␈α⊂to␈α⊂refer␈α⊃to␈α⊂data
␈↓ α←␈↓␈↓ β?structure␈α!behavior␈α"without␈α!regard␈α"to␈α!details␈α"of␈α!the
␈↓ α←␈↓␈↓ β?implementation␈α≥of␈α≥that␈α≥behavior␈α≥(``abstract␈α≥data␈α≤types''
␈↓ α←␈↓␈↓ β?[Liskov74]).

␈↓"β␈↓ α←␈↓␈↓ ββ(4)␈↓ β?␈↓↓Decomposition␈↓.  By␈α∪choosing␈α∪the␈α∪proper␈α∪boundaries,␈α∪systems
␈↓ α←␈↓␈↓ β?can␈α∃be␈α∀decomposed␈α∃into␈α∀modules␈α∃that␈α∀are␈α∃insulated␈α∀from
␈↓ α←␈↓␈↓ β?changes␈α∀in␈α∀one␈α∀another␈α∀(``information␈α∃hiding''␈α∀[Parnas72]).
␈↓ α←␈↓␈↓ β?(Consider␈αthe␈αdiscussion␈αabout␈α``traffic␈αdirectors''␈αin␈αSection␈α6-
␈↓ α←␈↓␈↓ β?9-1 in these terms.)

␈↓"β␈↓ α←␈↓␈↓ ββ(5)␈↓ β?␈↓↓Content-directed␈α#invocation␈↓.  By␈α#describing␈α#(rather␈α#than
␈↓ α←␈↓␈↓ β?naming)␈α→knowledge␈α_sources␈α→(KSs)␈α_and␈α→by␈α→effecting␈α_that
␈↓ α←␈↓␈↓ β?reference␈α~via␈α→direct␈α~examination␈α→of␈α~KS␈α~content,␈α→system
␈↓ α←␈↓␈↓ β?behavior␈α∞is␈α∞insulated␈α∞from␈α∞changes␈α∞to␈α∞the␈α∞code␈α∞of␈α∞any␈α∞of␈α
its
␈↓ α←␈↓␈↓ β?KSs.␈α
 Thus␈αa␈α
KS␈αcan␈α
be␈αedited␈α
and␈αits␈α
pattern␈α
of␈αinvocation
␈↓ α←␈↓␈↓ β?will␈α↔change␈α↔accordingly,␈α⊗without␈α↔the␈α↔necessity␈α↔of␈α⊗making
␈↓ α←␈↓␈↓ β?changes␈αelsewhere␈α
in␈αthe␈α
system␈α(see␈α
the␈αdiscussion␈α
in␈αSection
␈↓ α←␈↓␈↓ β?7-5-3).

␈↓"β␈↓ α←␈↓␈↓ β?There␈α
may␈α
quite␈α
plausibly,␈αhowever,␈α
be␈α
problems␈α
with␈α
elements␈αthat
␈↓ α←␈↓are␈α
inherently␈α
tightly␈α∞connected␈α
and␈α
complex.␈α
 It␈α∞may␈α
be␈α
impossible␈α∞to␈α
deal
␈↓ α←␈↓with␈αthese␈αwithout␈α
a␈αcorresponding␈αdegree␈αof␈α
complexity␈αin␈αthe␈αprogram␈α
that
␈↓ α←␈↓solves␈αthem.␈α Rather␈αthan␈αattempting␈αto␈αavoid␈αcomplexity,␈αtherefore,␈α
can␈αwe,
␈↓ α←␈↓in␈α
this␈α
case,␈α
design␈α
systems␈α
which␈α
are␈α
␈↓↓self-adjusting␈↓?␈α
 That␈α
is,␈α
is␈α
it␈αpossible␈α
to
␈↓ α←␈↓design␈αa␈αsystem␈αto␈αbe␈αinherently␈αflexible,␈αrather␈αthan␈αas␈αdrastically␈αfragile␈αas
␈↓ α←␈↓most programs typically are?
␈↓"β␈↓ α←␈↓␈↓ β?Achieving␈α∞such␈α∞flexibility␈α∞would␈α∞require␈α∞systems␈α∞that␈α∞are␈α
inherently
␈↓ α←␈↓stable,␈αin␈αwhich␈αthe␈αeffects␈αof␈αchanges␈αare␈αeasily␈αaccommodated.␈α Many␈αlarge
␈↓ α←␈↓programs␈α∪are␈α∪currently␈α∪similar␈α∪to␈α∪a␈α∪house␈α∪of␈α∪cards,␈α∪in␈α∪that␈α∪changes␈α∪or
␈↓ α←␈↓additions␈α∞must␈α∂be␈α∞made␈α∞with␈α∂utmost␈α∞care,␈α∞lest␈α∂the␈α∞entire␈α∂structure␈α∞collapse.
␈↓ α←␈↓There␈α(are␈α)simply␈α(too␈α(many␈α)unknown,␈α(unstable,␈α)and␈α(critical
␈↓ α←␈↓interdependencies.␈α What␈αwe␈αwould␈αlike␈αis␈αsomething␈αmore␈αakin␈αto␈αa␈αpond--
␈↓ α←␈↓␈↓274    CONCLUSIONS␈↓ 
#8-4␈↓

␈↓"β␈↓ α←␈↓throw␈αin␈αa␈αrock␈αor␈αadd␈αa␈αlot␈αof␈αwater␈αand␈αthere␈αis␈αa␈αcommotion␈αinitially,␈αbut
␈↓ α←␈↓eventually,␈αthe␈αripples␈αsubside,␈αthe␈α``system''␈αadjusts␈αto␈αthe␈αnew␈αconditions␈αby
␈↓ α←␈↓expanding␈α∩or␈α⊃reconfiguring␈α∩itself␈α⊃and␈α∩remains␈α⊃as␈α∩``competent''␈α⊃a␈α∩pond␈α⊃as
␈↓ α←␈↓before.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∞established␈α∞programming␈α∞technique␈α∞that␈α∞confronts␈α∂the␈α∞problem
␈↓ α←␈↓of␈α∃adjusting␈α∃to␈α∃changes␈α∃is␈α⊗the␈α∃idea␈α∃of␈α∃a␈α∃␈↓↓transform␈↓.␈α∃ It␈α⊗expresses␈α∃and
␈↓ α←␈↓automatically␈α⊂takes␈α⊂care␈α⊂of␈α⊂necessary␈α⊂interrelationships␈α⊂by␈α⊂propagating␈α∂the
␈↓ α←␈↓effects␈αof␈αchanges.␈α Consider␈αa␈αcompiler␈αin␈αthis␈αlight.␈α To␈αsee␈αthe␈αpoint␈αmost
␈↓ α←␈↓clearly,␈αimagine␈αin␈αits␈αabsence␈αthe␈αdifficulty␈αof␈αmaintaining␈αboth␈αsource␈αcode
␈↓ α←␈↓and␈α
machine␈αcode␈α
for␈αa␈α
program.␈α After␈α
any␈αediting␈α
in␈αthe␈α
source,␈α
it␈αwould
␈↓ α←␈↓still␈α∪be␈α∀necessary␈α∪to␈α∪edit␈α∀the␈α∪machine␈α∪code␈α∀analogously.␈α∪ But␈α∀given␈α∪the
␈↓ α←␈↓subtlety,␈αcomplexity,␈αand␈αamount␈αof␈αwork␈αinvolved␈αin␈αjust␈αunderstanding␈αthe
␈↓ α←␈↓mapping␈αfrom␈αsource␈αto␈αmachine␈αlanguage,␈αthe␈αtransform␈αhas␈αbeen␈αspecified
␈↓ α←␈↓precisely and turned over to the system.
␈↓"β␈↓ α←␈↓␈↓ β?An␈α
analogous␈α∞situation␈α
exists␈α
with␈α∞respect␈α
to␈α
data␈α∞structures.␈α
 There
␈↓ α←␈↓are␈α
programs␈α
with␈α
established,␈αcomplex␈α
interrelationships␈α
of␈α
data␈αstructures,
␈↓ α←␈↓and,␈α⊂at␈α∂very␈α⊂best,␈α∂the␈α⊂relationships␈α∂are␈α⊂only␈α∂documented␈α⊂somewhere.␈α∂ The
␈↓ α←␈↓programmer␈α⊃still␈α∩does␈α⊃multiple␈α∩edits␈α⊃of␈α⊃data␈α∩structures␈α⊃to␈α∩maintain␈α⊃those
␈↓ α←␈↓interrelationships,␈α⊃when␈α⊂he␈α⊃should␈α⊃be␈α⊂able␈α⊃to␈α⊃edit␈α⊂just␈α⊃one␈α⊃structure␈α⊂and
␈↓ α←␈↓``recompile␈αthe␈αdata␈αstructures''␈αto␈αhave␈αthe␈αsystem␈αtake␈αcare␈αof␈αthe␈α``ripples.'' 
␈↓ α←␈↓It␈α∞is␈α∞true␈α∞that␈α∞compilers␈α∞are␈α
both␈α∞better␈α∞understood␈α∞and␈α∞far␈α∞more␈α∞stable␈α
as
␈↓ α←␈↓transforms.␈α∩ Data␈α∩structure␈α∩interrelationships␈α⊃are␈α∩both␈α∩invented␈α∩anew␈α⊃for
␈↓ α←␈↓each␈αprogram␈α
and␈αsubject␈α
to␈αconstant␈α
change␈αas␈α
programs␈αevolve.␈α
 Even␈αso,
␈↓ α←␈↓we␈α⊂claim␈α⊂that␈α⊂there␈α⊂are␈α⊂interesting␈α⊂and␈α⊂important␈α⊂results␈α⊂that␈α⊃may␈α⊂follow
␈↓ α←␈↓from␈α
specifying␈αthe␈α
interrelationships.␈α
 First,␈αit␈α
can␈α
conceivably␈αmake␈α
possible
␈↓ α←␈↓larger␈αsystems␈αthan␈αmight␈αotherwise␈αbe␈αmaintainable,␈αand␈αsecond,␈αthe␈αtask␈αof
␈↓ α←␈↓being␈αprecise␈α
about␈αthose␈α
relationships␈αcan␈α
help␈αinsure␈α
that␈αthe␈α
design␈αis␈α
free
␈↓ α←␈↓from obvious inconsistencies.
␈↓"β␈↓ α←␈↓␈↓ β?Our␈α∞system␈α∞has␈α∂something␈α∞of␈α∞this␈α∂sort␈α∞in␈α∞the␈α∂␈↓∧RELATIONS␈↓␈α∞associated
␈↓ α←␈↓with␈α∞each␈α∞schema.␈α∞ They␈α∞specify␈α∞data␈α∞structure␈α∞interrelationships␈α∞and␈α∞offer
␈↓ α←␈↓one␈α
level␈α
of␈α
self-adjusting␈α
character: ␈α
The␈α
effects␈α
of␈α
adding␈α
a␈α
new␈αinstance␈α
of
␈↓ α←␈↓a␈αschema␈α
are␈αpropagated␈α
through␈αthe␈αrest␈α
of␈αthe␈α
system.␈α Recall,␈αfor␈α
example,
␈↓ α←␈↓the␈α∀extensive␈α∪interaction␈α∀triggered␈α∀by␈α∪adding␈α∀a␈α∀new␈α∪culture␈α∀site␈α∀to␈α∪the
␈↓ α←␈↓knowledge␈α
base␈αof␈α
the␈α
performance␈αprogram.␈α
 To␈α
push␈αthis␈α
one␈αlevel␈α
further,
␈↓ α←␈↓however,␈αconsider␈αa␈αmore␈αfundamental␈αchange.␈α What␈αif␈αit␈αwere␈αnecessary␈αto
␈↓ α←␈↓change␈αthe␈αdefinition␈αof␈αa␈αdata␈αstructure?␈α Part␈αof␈αthis␈αmight␈αinclude␈αwriting
␈↓ α←␈↓a␈αwhole␈αnew␈αset␈αof␈α␈↓∧RELATIONS␈↓␈αspecifications␈αfor␈αit␈αand␈αmaking␈α
modifications
␈↓ α←␈↓to the ␈↓∧RELATIONS␈↓ for other structures.
␈↓"β␈↓ α←␈↓␈↓ β?But␈αis␈αthere␈α
an␈αeasier␈αway␈α
to␈αdo␈αthis?␈α
 Would␈αit␈αbe␈α
possible␈αto␈αwrite␈α
a
␈↓ α←␈↓definition␈α
of␈α∞the␈α
data␈α∞structure␈α
and␈α
``compile''␈α∞it␈α
into␈α∞the␈α
appropriate␈α∞set␈α
of
␈↓ α←␈↓␈↓∧RELATIONS␈↓.␈α_ Recall,␈α_for␈α_example,␈α_the␈α_three␈α_categories␈α_of␈α→culture␈α_sites
␈↓ α←␈↓mentioned␈αin␈αchapter␈α
6.␈α It␈αshould␈αbe␈α
possible␈αto␈αsay␈α
that␈αevery␈αculture␈αsite␈α
is
␈↓ α←␈↓either␈α∀␈↓↓sterile,␈α∀non-sterile␈↓,␈α∀or␈α∀␈↓↓indeterminant␈↓,␈α∀that␈α∀is,␈α∀that␈α∀there␈α∃are␈α∀three
␈↓ α←␈↓mutually␈α∩exclusive␈α∩and␈α⊃exhaustive␈α∩classifications␈α∩for␈α⊃it.␈α∩ If,␈α∩later,␈α∩it␈α⊃were
␈↓ α←␈↓discovered␈α→that␈α→yet␈α~another␈α→classification␈α→were␈α→necessary,␈α~rather␈α→than
␈↓ α←␈↓␈↓8-4␈↓ ¬yTHE OTHER THEMES; SOME SPECULATIONS    275␈↓

␈↓"β␈↓ α←␈↓rewriting␈α∞all␈α
the␈α∞␈↓∧RELATIONS␈↓␈α∞by␈α
hand,␈α∞it␈α
should␈α∞be␈α∞possible␈α
to␈α∞edit␈α∞just␈α
the
␈↓ α←␈↓definition␈α∞of␈α∞a␈α∞␈↓∧SITE␈↓,␈α∞and␈α∂have␈α∞the␈α∞``data␈α∞structure␈α∞compiler''␈α∂recompute␈α∞all
␈↓ α←␈↓the new ␈↓∧RELATIONS␈↓.
␈↓"β␈↓ α←␈↓␈↓ β?This␈αwould␈αbe␈αa␈αuseful␈αcapability,␈αsince␈αrepresentation␈αredesign␈αis␈αan
␈↓ α←␈↓occupational␈αhazard␈αof␈αprogramming␈αand␈αone␈αthat␈αtypically␈αhas␈α
far-reaching
␈↓ α←␈↓effects␈αfor␈αeven␈αsmall␈αmodifications.␈α Logical␈αerrors␈αin␈αdesign␈αalso␈αoccur,␈αand
␈↓ α←␈↓with␈αmore␈αformal␈αdefinitions␈αit␈αmight␈αbe␈αpossible␈αto␈αcheck␈αfor␈αinconsistencies
␈↓ α←␈↓during the ``compilation.''
␈↓"β␈↓ α←␈↓␈↓ β?Note␈αthat␈αit␈αalso␈αextends␈αthe␈αconcept␈αof␈αa␈αself-adjusting␈αsystem.␈α One
␈↓ α←␈↓level␈αof␈αchange␈αpropagation␈αwas␈αaccomplished␈αby␈αthe␈αprecise␈αspecification␈αof
␈↓ α←␈↓the␈α
data␈α
structure␈α
interrelationships.␈α Another␈α
level␈α
would␈α
be␈α
made␈αpossible
␈↓ α←␈↓by␈α∀deriving␈α∃the␈α∀interrelationships␈α∀themselves␈α∃from␈α∀more␈α∃basic␈α∀elements,
␈↓ α←␈↓making it possible to propagate the effects of changes in definitions.
␈↓"β␈↓ α←␈↓␈↓ β?But␈α⊃all␈α⊃this␈α⊃is␈α⊃only␈α⊃speculation␈α⊃that␈α⊃raises␈α⊃more␈α⊃questions␈α⊃than␈α⊃it
␈↓ α←␈↓answers.␈α→ What␈α_constitutes␈α→a␈α_``definition''␈α→of␈α_a␈α→data␈α→structure?␈α_ What
␈↓ α←␈↓information␈α⊃should␈α⊃it␈α⊃contain?␈α⊃ The␈α⊃discussion␈α⊃above␈α⊃indicates␈α∩that␈α⊃both
␈↓ α←␈↓structural␈α
and␈α
behavioral␈α
definitions␈α
have␈α
been␈α
proposed;␈α
there␈α
may␈α∞be␈α
yet
␈↓ α←␈↓other␈α∞possibilities.␈α∞ How␈α∞is␈α∞the␈α∞information␈α∞to␈α∞be␈α∞represented?␈α∂ Much␈α∞work
␈↓ α←␈↓has␈αbeen␈αdirected␈αtoward␈αthe␈αuse␈αof␈αformal␈αlanguages,␈αand␈αhas␈αdemonstrated
␈↓ α←␈↓the␈α∩complexity␈α∩that␈α∩lies␈α∪beneath␈α∩apparently␈α∩simple␈α∩structures␈α∪when␈α∩their
␈↓ α←␈↓properties are described precisely (see, for instance, [Spitzen75]).
␈↓"β␈↓ α←␈↓␈↓ β?One␈α∞further␈α∞speculation␈α∂to␈α∞set␈α∞all␈α∞this␈α∂in␈α∞the␈α∞broadest␈α∂terms: ␈α∞Some
␈↓ α←␈↓elements␈αof␈α
the␈αtechniques␈αthat␈α
have␈αbeen␈αemployed␈α
in␈αdealing␈α
with␈αchange
␈↓ α←␈↓seem␈α
general␈α∞enough␈α
to␈α∞be␈α
widely␈α
applicable.␈α∞ Can␈α
this␈α∞set␈α
of␈α∞principles␈α
be
␈↓ α←␈↓extended␈α
and␈α∞made␈α
large␈α∞enough␈α
to␈α
provide␈α∞useful␈α
guidance?␈α∞ Might␈α
there
␈↓ α←␈↓someday be a ``science'' of representation design?

␈↓"β␈↓ α←␈↓␈↓α8-5    PROJECTIONS␈↓
␈↓"β␈↓ α←␈↓␈↓ β?The␈αpast␈αseven␈αchapters␈αhave␈αexplored␈αsome␈αpreliminary␈αsolutions␈αto
␈↓ α←␈↓the␈α_myriad␈α_of␈α→problems␈α_encountered␈α_in␈α→the␈α_attempt␈α_to␈α→facilitate␈α_the
␈↓ α←␈↓construction␈α
of␈α
expert␈α
systems.␈α
 This␈α
is␈α
only␈α
a␈α
beginning,␈α
for␈α
much␈αremains
␈↓ α←␈↓to␈αbe␈αdone.␈α But␈αwhat␈αmight␈αbe␈αthe␈αeventual␈αimpact␈αof␈αsuccess?␈α What␈αmight
␈↓ α←␈↓be␈α
gained␈αby␈α
solving␈αthe␈α
problems?␈α Two␈α
simple␈αbut␈α
significant␈αpossibilities
␈↓ α←␈↓come to mind.
␈↓"β␈↓ α←␈↓␈↓ β?One␈α
result␈α
might␈αbe␈α
the␈α
realization␈α
of␈αa␈α
different␈α
form␈α
of␈αgenerality
␈↓ α←␈↓in␈αproblem␈αsolving.␈α This␈αreport␈αbegan␈αwith␈αthe␈αobservation␈αthat␈α
the␈αsearch
␈↓ α←␈↓for␈αgenerality␈α
in␈αAI␈α
programs␈αhas␈α
traditionally␈αfocused␈α
on␈αbroadly␈α
applicable
␈↓ α←␈↓problem-solving␈α
methods␈α
and␈α
has␈α
so␈α
far␈α
proved␈α
unsuccessful.␈α
 An␈α
alternative
␈↓ α←␈↓form␈αof␈αgenerality␈αwas␈αsuggested,␈αone␈αthat␈αtook␈αits␈αbreadth␈αnot␈αfrom␈αgeneral
␈↓ α←␈↓problem-solving␈α methods,␈α∨but␈α from␈α∨powerful␈α knowledge␈α∨acquisition
␈↓ α←␈↓techniques.

␈↓"β␈↓ α←␈↓␈↓ β'The␈αappropriate␈αplace␈αfor␈αan␈αattack␈αon␈αthe␈αproblem␈αof␈αgenerality
␈↓ α←␈↓␈↓ β'may␈α
be␈α
at␈α
the␈αmeta-levels␈α
of␈α
learning,␈α
knowledge␈αtransformation,
␈↓ α←␈↓␈↓ β'and␈α∪representation,␈α∩not␈α∪at␈α∪the␈α∩level␈α∪of␈α∪performance␈α∩programs.
␈↓ α←␈↓␈↓276    CONCLUSIONS␈↓ 
#8-5␈↓

␈↓"β␈↓ α←␈↓␈↓ β'Perhaps␈α∃for␈α∀the␈α∃designer␈α∀of␈α∃intelligent␈α∀systems␈α∃what␈α∃is␈α∀most
␈↓ α←␈↓␈↓ β'significant␈α∞about␈α∞human␈α∞problem␈α∞solving␈α∞behavior␈α∞is␈α∞the␈α
ability
␈↓ α←␈↓␈↓ β'to␈αlearn␈αspecialties␈αas␈αneeded--to␈αlearn␈αexpertise␈αin␈αproblem␈αareas
␈↓ α←␈↓␈↓ β'by␈α∪learning␈α∪problem-specific␈α∩heuristics,␈α∪by␈α∪acquiring␈α∩problem-
␈↓ α←␈↓␈↓ β'specific␈αinformation,␈α
and␈αby␈α
transforming␈αgeneral␈α
knowledge␈αand
␈↓ α←␈↓␈↓ β'general processes into specialized forms.
␈↓"β␈↓ α←␈↓␈↓ βo␈↓ λ7[Feigenbaum71] 

␈↓ α←␈↓We␈αhave␈αtaken␈αa␈αstep␈αin␈αthis␈αdirection␈αby␈αmaking␈αpossible␈αa␈αlimited␈αform␈αof
␈↓ α←␈↓communication␈α⊂between␈α⊂an␈α⊂expert␈α⊂in␈α⊂an␈α⊂application␈α⊂domain␈α⊂and␈α⊃a␈α⊂high-
␈↓ α←␈↓performance␈α∩program.␈α∩ The␈α∩program␈α∩is␈α∩capable␈α∩of␈α∩explaining␈α⊃significant
␈↓ α←␈↓portions␈α∪of␈α∪its␈α∪behavior␈α∪and␈α∪is␈α∪prepared␈α∪to␈α∪accept␈α∪new␈α∀knowledge␈α∪and
␈↓ α←␈↓integrate it into its knowledge base.
␈↓"β␈↓ α←␈↓␈↓ β?Finally,␈α∀if␈α∃it␈α∀were␈α∃truly␈α∀easy␈α∃to␈α∀construct␈α∃large␈α∀knowledge-based
␈↓ α←␈↓systems,␈α↔then␈α↔might␈α↔they␈α_eventually␈α↔become␈α↔an␈α↔important␈α_medium␈α↔of
␈↓ α←␈↓expression?␈α→ Might␈α→they␈α→provide␈α_a␈α→medium␈α→for␈α→the␈α→formalization␈α_of
␈↓ α←␈↓knowledge␈α⊂in␈α⊂domains␈α⊂where␈α⊂it␈α⊂is␈α⊂as␈α⊂yet␈α⊂highly␈α⊂informal,␈α⊂and␈α⊂could␈α∂they
␈↓ α←␈↓provide␈α∀a␈α∀useful␈α∀framework␈α∀for␈α∀its␈α∀expression␈α∀and␈α∀organization?␈α∀ Such
␈↓ α←␈↓systems␈αmight␈α
help␈αto␈α
collect␈αand␈α
organize␈αknowledge␈α
on␈αa␈α
scale␈αlarge␈α
enough
␈↓ α←␈↓to␈αbecome␈α
useful␈αtools␈α
in␈αthe␈α
attempts␈αto␈α
understand␈αand␈α
develop␈αtheories␈α
for
␈↓ α←␈↓domains where no cohesive theories yet exist.
␈↓ α←␈↓␈↓8-5␈↓ 
α    277␈↓























␈↓"β␈↓ α←␈↓␈↓ βWI will say nothing further.
␈↓ α←␈↓␈↓ βWAgainst this answer let your temper rage as wildly as you will.

␈↓"β␈↓ α←␈↓␈↓ βW␈↓ 	#lines 343-344 
␈↓ α←␈↓␈↓278␈↓ 	↔AUTHOR INDEX␈↓


␈↓"β␈↓ α←␈↓α␈↓ ¬{Author Index



␈↓"β␈↓ α←␈↓[Aristotle26]␈↓ ∧[42                              ␈↓ ε{␈↓[Harre70]␈↓ λw77
                                             ␈↓ ε{␈↓[Hart75]␈↓ λw1, 12
␈↓ α←␈↓[Bachman75]␈↓ ∧[131, 273                          ␈↓ ε{␈↓[Hayes-Roth76]␈↓ λw85
␈↓ α←␈↓[Balzer67]␈↓ ∧[131, 185, 273                      ␈↓ ε{␈↓[Hewitt71]␈↓ λw52
␈↓ α←␈↓[Barrow75]␈↓ ∧[77                                 ␈↓ ε{␈↓[Hewitt75]␈↓ λw157
␈↓ α←␈↓[Baumgart74]␈↓ ∧[78                               ␈↓ ε{␈↓[Howe73]␈↓ λw242
␈↓ α←␈↓[Bobrow68]␈↓ ∧[196
␈↓ α←␈↓[Bobrow77]␈↓ ∧[227                                ␈↓ ε{␈↓[Interaction72]␈↓ λw241
␈↓ α←␈↓[Brachman75]␈↓ ∧[60
␈↓ α←␈↓[Brown75]␈↓ ∧[50, 51                              ␈↓ ε{␈↓[Johnson75]␈↓ λw61
␈↓ α←␈↓[Buchanan71]␈↓ ∧[1
␈↓ α←␈↓[Buchanan72]␈↓ ∧[60                               ␈↓ ε{␈↓[Kulikowski73]␈↓ λw12, 50

␈↓ α←␈↓[Carbonell73]␈↓ ∧[111, 112                        ␈↓ ε{␈↓[Learning76]␈↓ λw157
␈↓ α←␈↓[Carroll60]␈↓ ∧[31                                ␈↓ ε{␈↓[Lesser74]␈↓ λw27
                                             ␈↓ ε{␈↓[Liskov74]␈↓ λw131, 273
␈↓ α←␈↓[Dahl70]␈↓ ∧[145                                  ␈↓ ε{␈↓[Low74]␈↓ λw237
␈↓ α←␈↓[Davis77a]␈↓ ∧[22, 22, 48, 217,                   ␈↓ ε{␈↓[Lukasciewicz70]␈↓ λw16
␈↓ α←␈↓␈↓ β∂222
␈↓ α←␈↓[Davis77b]␈↓ ∧[12                                 ␈↓ ε{␈↓[MACSYMA74]␈↓ λw1
␈↓ α←␈↓[Dijkstra72]␈↓ ∧[37, 272                          ␈↓ ε{␈↓[Manna69]␈↓ λw264
                                             ␈↓ ε{␈↓[McDermott74]␈↓ λw60
␈↓ α←␈↓[Eswarn75]␈↓ ∧[134                                ␈↓ ε{␈↓[McLeod76]␈↓ λw61, 134
                                             ␈↓ ε{␈↓[Miller75]␈↓ λw216, 238
␈↓ α←␈↓[Falk70]␈↓ ∧[78, 79, 114, 115                     ␈↓ ε{␈↓[Minksy68]␈↓ λw269
␈↓ α←␈↓[Faught74]␈↓ ∧[51                                 ␈↓ ε{␈↓[Minsky74]␈↓ λw146
␈↓ α←␈↓[Feigenbaum71]␈↓ ∧[5, 276                         ␈↓ ε{␈↓[Mitchell70]␈↓ λw237
␈↓ α←␈↓[Feldman72]␈↓ ∧[222
␈↓ α←␈↓[Fikes71]␈↓ ∧[220                                 ␈↓ ε{␈↓[Newell59]␈↓ λw223
␈↓ α←␈↓[Fikes72]␈↓ ∧[245                                 ␈↓ ε{␈↓[Newell61]␈↓ λw223
␈↓ α←␈↓[Finkel74]␈↓ ∧[1                                  ␈↓ ε{␈↓[Newell69]␈↓ λw196, 198, 249
                                             ␈↓ ε{␈↓[Newell72]␈↓ λw220
␈↓ α←␈↓[Gelernter59]␈↓ ∧[238, 242                        ␈↓ ε{␈↓[Norman75]␈↓ λw79
␈↓ α←␈↓[Goldstein74]␈↓ ∧[52, 61, 114
␈↓ α←␈↓[Green69]␈↓ ∧[238, 238, 241,                      ␈↓ ε{␈↓[Parnas72]␈↓ λw273
␈↓ α←␈↓␈↓ β∂243                                          ␈↓ ε{␈↓[Parnas75]␈↓ λw270
␈↓ α←␈↓[Green74]␈↓ ∧[264                                 ␈↓ ε{␈↓[Polya54]␈↓ λw241
␈↓ α←␈↓[Gregory66]␈↓ ∧[77                                ␈↓ ε{␈↓[Pople75]␈↓ λw12
␈↓ α←␈↓[Guzman68]␈↓ ∧[78                                 ␈↓ ε{␈↓[Post43]␈↓ λw22

␈↓ α←␈↓[Hansen74]␈↓ ∧[237                                ␈↓ ε{␈↓[Reddy73]␈↓ λw114, 114
␈↓ α←␈↓␈↓AUTHOR INDEX␈↓ 
"279␈↓






␈↓"β␈↓ α←␈↓[Reiger74]␈↓ ∧[51
␈↓ α←␈↓[Roberts63]␈↓ ∧[78
␈↓ α←␈↓[Rubin75]␈↓ ∧[238
␈↓ α←␈↓[Rulifson72]␈↓ ∧[132, 222
␈↓ α←␈↓[Rumelhart73]␈↓ ∧[98

␈↓ α←␈↓[Sacerdoti73]␈↓ ∧[242, 243
␈↓ α←␈↓[Sacerdoti77]␈↓ ∧[50
␈↓ α←␈↓[Samet75]␈↓ ∧[237
␈↓ α←␈↓[Samuel67]␈↓ ∧[245
␈↓ α←␈↓[Sandewall75]␈↓ ∧[61, 189
␈↓ α←␈↓[Shaw75]␈↓ ∧[263
␈↓ α←␈↓[Shortliffe75a]␈↓ ∧[52
␈↓ α←␈↓[Shortliffe75b]␈↓ ∧[16
␈↓ α←␈↓[Shortliffe76]␈↓ ∧[12, 45, 98, 120,
␈↓ α←␈↓␈↓ β∂151
␈↓ α←␈↓[Simon73]␈↓ ∧[196
␈↓ α←␈↓[Sophocles27]␈↓ ∧[1
␈↓ α←␈↓[Spitzen75]␈↓ ∧[186, 263, 275
␈↓ α←␈↓[Stonebreaker75]␈↓ ∧[186
␈↓ α←␈↓[Sussman75]␈↓ ∧[61
␈↓ α←␈↓[Suzuki76]␈↓ ∧[186, 263, 266

␈↓ α←␈↓[Tversky74]␈↓ ∧[33

␈↓ α←␈↓[vanMelle74]␈↓ ∧[21

␈↓ α←␈↓[Waldinger74]␈↓ ∧[227, 245, 264
␈↓ α←␈↓[Waltz72]␈↓ ∧[114, 114, 115
␈↓ α←␈↓[Warshall62]␈↓ ∧[208
␈↓ α←␈↓[Waterman70]␈↓ ∧[23
␈↓ α←␈↓[Waterman77]␈↓ ∧[60, 122
␈↓ α←␈↓[Wickelgren74]␈↓ ∧[241
␈↓ α←␈↓[Winograd72]␈↓ ∧[114
␈↓ α←␈↓[Winograd74]␈↓ ∧[271
␈↓ α←␈↓[Winograd75]␈↓ ∧[249
␈↓ α←␈↓[Winston70]␈↓ ∧[59, 84, 85, 115
␈↓ α←␈↓␈↓280␈↓ 	3TOPIC INDEX␈↓


␈↓"β␈↓ α←␈↓α␈↓ εβTopic Index



␈↓"β␈↓ α←␈↓abstraction  49                              ␈↓ ε{␈↓flexibility  3, 172, 200, 219, 229,
␈↓ α←␈↓associative triple  17, 21                   ␈↓ ε{␈↓␈↓ π+236, 238, 239, 240, 248, 258
                                             ␈↓ ε{␈↓flexible  27
␈↓ α←␈↓backtracking  99                             ␈↓ ε{␈↓focus, in explanation  46
␈↓ α←␈↓backward chaining  18, 24, 34                ␈↓ ε{␈↓focus, in explanations  90
␈↓ α←␈↓body of a KS  220                            ␈↓ ε{␈↓focus, in knowledge acquisition  61,
␈↓ α←␈↓bootstrapping the knowledge base             ␈↓ ε{␈↓␈↓ π+69
␈↓ α←␈↓␈↓ β∂182                                          ␈↓ ε{␈↓focusing  89

␈↓ α←␈↓complexity  237, 271, 272                    ␈↓ ε{␈↓generality  29
␈↓ α←␈↓computer consultants  11                     ␈↓ ε{␈↓generalization  47
␈↓ α←␈↓computer vision  77                          ␈↓ ε{␈↓generalization hierarchy  143
␈↓ α←␈↓concept formation  84, 113                   ␈↓ ε{␈↓generalized invocation criteria  219,
␈↓ α←␈↓conflict set  197                            ␈↓ ε{␈↓␈↓ π+258
␈↓ α←␈↓consistency  71, 94, 100, 103, 119,          ␈↓ ε{␈↓generator  162
␈↓ α←␈↓␈↓ β∂120                                          ␈↓ ε{␈↓goal tree  18, 36, 58
␈↓ α←␈↓content-directed invocation  195,
␈↓ α←␈↓␈↓ β∂219-231, 258                                 ␈↓ ε{␈↓handle of a KS  220
␈↓ α←␈↓contradiction  120                           ␈↓ ε{␈↓high-level description  87
                                             ␈↓ ε{␈↓high-level language  25, 34
␈↓ α←␈↓data structure interrelationships            ␈↓ ε{␈↓human engineering  35, 56, 89, 155
␈↓ α←␈↓␈↓ β∂132
␈↓ α←␈↓data types  25                               ␈↓ ε{␈↓implicit context  22
␈↓ α←␈↓demon  158                                   ␈↓ ε{␈↓inference engine  4, 15, 18-20
␈↓ α←␈↓description  78, 87                          ␈↓ ε{␈↓inference rules  6, 16, 18, 21
␈↓ α←␈↓distribution of knowledge  157, 159          ␈↓ ε{␈↓information metric  42, 49
␈↓ α←␈↓documentation  138, 150                      ␈↓ ε{␈↓inheritance of properties  144, 174
␈↓ α←␈↓domain independence  5, 20                   ␈↓ ε{␈↓insulation  163, 272
                                             ␈↓ ε{␈↓interactive transfer of expertise
␈↓ α←␈↓empirical generalizations  79                ␈↓ ε{␈↓␈↓ π+173, 255
␈↓ α←␈↓exhaustive invocation  193
␈↓ α←␈↓expectations  6, 61, 73, 77, 87, 92,         ␈↓ ε{␈↓knowledge about representations
␈↓ α←␈↓␈↓ β∂102, 255                                     ␈↓ ε{␈↓␈↓ π+126, 130, 173, 183, 256
␈↓ α←␈↓expertise  3, 11                             ␈↓ ε{␈↓knowledge acquisition  3, 6, 15, 28,
␈↓ α←␈↓explaining meta-rules  214                   ␈↓ ε{␈↓␈↓ π+29
␈↓ α←␈↓explanation  3, 6, 15                        ␈↓ ε{␈↓knowledge acquisition in context
␈↓ α←␈↓expressiveness  219, 258                     ␈↓ ε{␈↓␈↓ π+73, 77, 106, 107, 122
␈↓ α←␈↓extended data type  52, 98, 126,             ␈↓ ε{␈↓knowledge base  3, 4, 15, 16-18
␈↓ α←␈↓␈↓ β∂127, 130, 131                                ␈↓ ε{␈↓knowledge base integrity  134, 149
                                             ␈↓ ε{␈↓knowledge base maintenance  270
␈↓ α←␈↓␈↓TOPIC INDEX␈↓ 
"281␈↓






␈↓"β␈↓ α←␈↓knowledge base management  3, 6,             ␈↓ ε{␈↓smart tree  209
␈↓ α←␈↓␈↓ β∂61, 74, 76, 108, 127, 128                    ␈↓ ε{␈↓strategies  6, 15, 28
␈↓ α←␈↓knowledge organization  24, 163              ␈↓ ε{␈↓subgoal  18
␈↓ α←␈↓knowledge representation  23, 28,            ␈↓ ε{␈↓subsumption  120
␈↓ α←␈↓␈↓ β∂127, 130                                     ␈↓ ε{␈↓symbolic reasoning  32
␈↓ α←␈↓knowledge representations  3, 22
␈↓ α←␈↓knowledge sources, in rule                   ␈↓ ε{␈↓target  156
␈↓ α←␈↓␈↓ β∂acquisition  93                              ␈↓ ε{␈↓traffic director  157-160
␈↓ α←␈↓knowledge transfer  3                        ␈↓ ε{␈↓transform  274
                                             ␈↓ ε{␈↓tree of parses  98
␈↓ α←␈↓level of detail  32, 42                      ␈↓ ε{␈↓tree search  19
␈↓ α←␈↓levels of knowledge  183-186, 201            ␈↓ ε{␈↓tree traversal  36, 37, 58
␈↓ α←␈↓line of reasoning  211                       ␈↓ ε{␈↓trigger  156
                                             ␈↓ ε{␈↓type checking  133
␈↓ α←␈↓meta-level knowledge  6, 27, 86,
␈↓ α←␈↓␈↓ β∂131, 253, 271                                ␈↓ ε{␈↓unity path   19
␈↓ α←␈↓model of knowledge  73, 81, 87               ␈↓ ε{␈↓updating function  156-160
␈↓ α←␈↓model-based understanding  77,
␈↓ α←␈↓␈↓ β∂113-117                                      ␈↓ ε{␈↓validity  219, 258
␈↓ α←␈↓model-directed understanding  255            ␈↓ ε{␈↓verification  78, 87
␈↓ α←␈↓models  78, 79, 113-117

␈↓ α←␈↓partial evaluation  26, 46, 107
␈↓ α←␈↓partial evaluation   19
␈↓ α←␈↓partial match  105
␈↓ α←␈↓pattern matching  161, 168
␈↓ α←␈↓plausible knowledge source set  197
␈↓ α←␈↓principle line  245
␈↓ α←␈↓procedural encoding  24
␈↓ α←␈↓production rule  259
␈↓ α←␈↓production rules  4, 8, 22-27, 34, 48

␈↓ α←␈↓reasoning chain  37
␈↓ α←␈↓recognition  78, 87, 92, 102, 255

␈↓ α←␈↓saturation  194
␈↓ α←␈↓schema  127
␈↓ α←␈↓search algorithm  209
␈↓ α←␈↓second guess  259
␈↓ α←␈↓second guessing  73, 105, 119
␈↓ α←␈↓self-adjusting systems  273
␈↓ α←␈↓␈↓282␈↓ 	0REFERENCES␈↓

␈↓"β␈↓ α←␈↓␈↓ ε∞␈↓αReferences␈↓


␈↓"β␈↓ α←␈↓Several abbreviations are used for readability:

␈↓"β␈↓ α←␈↓3IJCAI ␈↓ ∧π␈↓↓Proceedings␈α⊗of␈α⊗the␈α∃Third␈α⊗International␈α⊗Joint␈α⊗Conference␈α∃on
␈↓ α←␈↓↓␈↓ ∧πArtificial␈α Intelligence␈↓,␈α [available␈α from␈α SRI␈α International,
␈↓ α←␈↓␈↓ ∧πPublications, 330 Ravenswood Ave, Menlo Park, CA 94025].

␈↓"β␈↓ α←␈↓4IJCAI ␈↓ ∧π␈↓↓Proceedings␈α∀of␈α∪the␈α∀Fourth␈α∪International␈α∀Joint␈α∀Conference␈α∪on
␈↓ α←␈↓↓␈↓ ∧πArtificial␈α_Intelligence␈↓,␈α_[available␈α→from␈α_MIT␈α_A.I.␈α→Lab,␈α_545
␈↓ α←␈↓␈↓ ∧πTechnology Square, Cambridge, MA 02138].

␈↓"β␈↓ α←␈↓AIM # ␈↓ ∧πA.␈α≤I.␈α≤Memo␈α≠#,␈α≤Computer␈α≤Science␈α≤Department,␈α≠Stanford,
␈↓ α←␈↓␈↓ ∧πCalifornia.

␈↓"β␈↓ α←␈↓MIT ␈↓ ∧πMassachusetts Institute of Technology, Cambridge, MA.

␈↓"β␈↓ α←␈↓CMU ␈↓ ∧πCarnegie Mellon University, Pittsburgh, Pennsylvania.



␈↓"β␈↓ α←␈↓[ARISTOTLE26]
␈↓"β␈↓ α←␈↓␈↓ β∂Aristotle: ␈↓↓Rhetoric␈↓, J. H. Freese (trans.), G. P. Putnam, New York, 1926.

␈↓"β␈↓ α←␈↓[Bachman75]
␈↓"β␈↓ α←␈↓␈↓ β∂Bachman,␈α∃C. W.:␈α∃Trends␈α⊗in␈α∃data␈α∃base␈α⊗management--1975,␈α∃␈↓↓National
␈↓ α←␈↓↓␈↓ β∂Computer Conference␈↓, 1975, pp. 569-576.

␈↓"β␈↓ α←␈↓[Balzer67]
␈↓"β␈↓ α←␈↓␈↓ β∂Balzer,␈αR. M.:␈α
Dataless␈αprogramming,␈α
␈↓↓Proc.␈αAFIPS␈α
Conf.␈↓,␈α1967,␈α
pp.␈α535-
␈↓ α←␈↓␈↓ β∂544.

␈↓"β␈↓ α←␈↓[Barrow75]
␈↓"β␈↓ α←␈↓␈↓ β∂Barrow,␈α⊗H. G.,␈α⊗and␈α⊗J. M.␈α⊗Tannenbaum:␈α⊗Representation␈α⊗and␈α⊗use␈α⊗of
␈↓ α←␈↓␈↓ β∂knowledge␈α
in␈αvision,␈α
␈↓↓SIGART␈α
Newsletter␈↓,␈αJune␈α
1975,␈αpp.␈α
2-8.␈α
 (Also␈αsee
␈↓ α←␈↓␈↓ β∂the articles following it.)

␈↓"β␈↓ α←␈↓[Baumgart74]
␈↓"β␈↓ α←␈↓␈↓ β∂Baumgart,␈α⊗B. G.:␈α⊗Geometric␈α⊗models␈α⊗for␈α⊗computer␈α⊗vision,␈α⊗AIM␈α⊗249,
␈↓ α←␈↓␈↓ β∂October 1974.

␈↓"β␈↓ α←␈↓[Bobrow68]
␈↓"β␈↓ α←␈↓␈↓ β∂Bobrow,␈α∞D. N.:␈α∞Natural␈α
language␈α∞input␈α∞for␈α
a␈α∞computer␈α∞problem␈α
solving
␈↓ α←␈↓␈↓ β∂system,␈α∂in␈α⊂M. Minsky␈α∂(ed.),␈α∂␈↓↓Semantic␈α⊂Information␈α∂Processing␈↓,␈α⊂The␈α∂MIT
␈↓ α←␈↓␈↓ β∂Press, MIT, 1968, pp. 146-226.
␈↓ α←␈↓␈↓REFERENCES␈↓ 
"283␈↓

␈↓"β␈↓ α←␈↓[Bobrow75]
␈↓"β␈↓ α←␈↓␈↓ β∂Bobrow,␈α
D. N.,␈α
and␈α
A. Collins␈α
(eds.):␈α
␈↓↓Representation␈α
and␈α
Understanding␈↓,
␈↓ α←␈↓␈↓ β∂Academic Press, New York, 1975.

␈↓"β␈↓ α←␈↓[Bobrow77]
␈↓"β␈↓ α←␈↓␈↓ β∂Bobrow,␈α
D. N.,␈α∞and␈α
T.␈α∞Winograd:␈α
Overview␈α∞of␈α
KRL,␈α∞␈↓↓Cognitive␈α
Science␈↓,
␈↓ α←␈↓␈↓ β∂vol. 1, January 1977, pp. 3-47.

␈↓"β␈↓ α←␈↓[Brachman75]
␈↓"β␈↓ α←␈↓␈↓ β∂Brachman,␈α→R. J.:␈α→Structural␈α_knowledge␈α→in␈α→a␈α→document␈α_information
␈↓ α←␈↓␈↓ β∂consulting␈α≤system,␈α≠TR␈α≤6-75,␈α≤Center␈α≠for␈α≤Research␈α≤in␈α≠Computing
␈↓ α←␈↓␈↓ β∂Technology, Harvard University, Cambridge, 1975.

␈↓"β␈↓ α←␈↓[Brown75]
␈↓"β␈↓ α←␈↓␈↓ β∂Brown,␈α
J. S.:␈αUses␈α
of␈αAI␈α
and␈α
advanced␈αcomputer␈α
technology␈αin␈α
education,
␈↓ α←␈↓␈↓ β∂Bolt, Beranek, & Newman, Cambridge, December 1975.

␈↓"β␈↓ α←␈↓[Buchanan71]
␈↓"β␈↓ α←␈↓␈↓ β∂Buchanan,␈α⊃B. G.,␈α⊂and␈α⊃J. Lederberg:␈α⊃The␈α⊂heuristic␈α⊃␈↓¬DENDRAL␈↓␈α⊃program␈α⊂for
␈↓ α←␈↓␈↓ β∂explaining empirical data, ␈↓↓IFIP␈↓, 1971, pp. 179-188.

␈↓"β␈↓ α←␈↓[Buchanan72]
␈↓"β␈↓ α←␈↓␈↓ β∂Buchanan,␈α∃B. G.,␈α⊗E. A.␈α∃Feigenbaum,␈α⊗and␈α∃N. S.␈α⊗Sridharan:␈α∃Heuristic
␈↓ α←␈↓␈↓ β∂theory␈α
formation: ␈α
Data␈α
interpretation␈α
and␈α
rule␈α
formation,␈α∞in␈α
B. Meltzer
␈↓ α←␈↓␈↓ β∂and␈αD. Michie␈α
(eds.),␈α␈↓↓Machine␈αIntelligence␈α
7␈↓,␈αEdinburgh␈αUniversity␈α
Press,
␈↓ α←␈↓␈↓ β∂Edinburgh, 1972, pp. 267-292.

␈↓"β␈↓ α←␈↓[Carbonell73]
␈↓"β␈↓ α←␈↓␈↓ β∂Carbonell,␈α⊂J. R.,␈α⊂and␈α⊂A. M.␈α⊃Collins:␈α⊂Natural␈α⊂semantics␈α⊂in␈α⊃AI,␈α⊂␈↓↓3IJCAI␈↓,
␈↓ α←␈↓␈↓ β∂1973, pp. 344-351.

␈↓"β␈↓ α←␈↓[Carroll60]
␈↓"β␈↓ α←␈↓␈↓ β∂Carroll,␈α
L.:␈αAlice's␈α
Adventures␈αin␈α
Wonderland,␈αin␈α
M. Gardner␈α
(ed.),␈α␈↓↓The
␈↓ α←␈↓↓␈↓ β∂Annotated Alice␈↓, World Publishing Company, 1960.

␈↓"β␈↓ α←␈↓[Dahl70]
␈↓"β␈↓ α←␈↓␈↓ β∂Dahl,␈α⊃O. J.,␈α∩B.␈α⊃Myhrhaug,␈α∩and␈α⊃K.␈α∩Nygaard:␈α⊃Common␈α∩base␈α⊃language,
␈↓ α←␈↓␈↓ β∂␈↓↓Norwegian Computing Center Technical Report␈↓, 1970.

␈↓"β␈↓ α←␈↓[Davis77a]
␈↓"β␈↓ α←␈↓␈↓ β∂Davis,␈α⊂R.,␈α⊃and␈α⊂J.␈α⊃King:␈α⊂An␈α⊃overview␈α⊂of␈α⊃production␈α⊂systems,␈α⊃in␈α⊂E. W.
␈↓ α←␈↓␈↓ β∂Elcock␈αand␈α
D. Michie␈α(eds.),␈α
␈↓↓Machine␈αIntelligence 8␈↓,␈α
John␈αWiley␈α
&␈αSons,
␈↓ α←␈↓␈↓ β∂New York, 1977  (also AIM 271).
␈↓ α←␈↓␈↓284␈↓ 	0REFERENCES␈↓

␈↓"β␈↓ α←␈↓[Davis77b]
␈↓"β␈↓ α←␈↓␈↓ β∂Davis,␈α∀R.,␈α∀B. G.␈α∀Buchanan,␈α∀and␈α∀E. Shortliffe:␈α∀Production␈α∀rules␈α∀as␈α∪a
␈↓ α←␈↓␈↓ β∂representation␈α∀for␈α∪a␈α∀knowledge-based␈α∪consultation␈α∀program,␈α∪␈↓↓Artificial
␈↓ α←␈↓↓␈↓ β∂Intelligence␈↓, vol. 8, no. 1, February 1977, pp. 15-45 (also AIM 266).

␈↓"β␈↓ α←␈↓[Dijkstra72]
␈↓"β␈↓ α←␈↓␈↓ β∂Dijkstra,␈α⊂E.,␈α⊂O. J.␈α⊂Dahl,␈α∂and␈α⊂C. A. R.␈α⊂Hoare:␈α⊂␈↓↓Structured␈α∂Programming␈↓,
␈↓ α←␈↓␈↓ β∂Academic Press, New York, 1972.

␈↓"β␈↓ α←␈↓[Eswaran75]
␈↓"β␈↓ α←␈↓␈↓ β∂Eswaran,␈α∩K. P.,␈α⊃and␈α∩D. D.␈α⊃Chamberlain:␈α∩Functional␈α⊃specification␈α∩of␈α⊃a
␈↓ α←␈↓␈↓ β∂subsystem␈αfor␈αdata␈αbase␈αintegrity,␈α␈↓↓Proceedings␈αof␈αInternational␈αConference
␈↓ α←␈↓↓␈↓ β∂on Very Large Data Bases␈↓, September 1975.

␈↓"β␈↓ α←␈↓[Falk70]
␈↓"β␈↓ α←␈↓␈↓ β∂Falk,␈α∃G.:␈α∃Computer␈α∀interpretation␈α∃of␈α∃imperfect␈α∀line␈α∃data,␈α∃AIM␈α∀132,
␈↓ α←␈↓␈↓ β∂August 1970.

␈↓"β␈↓ α←␈↓[Faught74]
␈↓"β␈↓ α←␈↓␈↓ β∂Faught,␈α∀W.,␈α∀K.␈α∃Colby,␈α∀and␈α∀R.␈α∃Parkison:␈α∀The␈α∀interaction␈α∃of␈α∀affects,
␈↓ α←␈↓␈↓ β∂intentions and desires, AIM 253, December 1974.

␈↓"β␈↓ α←␈↓[Feigenbaum71]
␈↓"β␈↓ α←␈↓␈↓ β∂Feigenbaum,␈α∂E. A.,␈α∂B.␈α∞Buchanan,␈α∂and␈α∂J.␈α∞Lederberg:␈α∂On␈α∂generality␈α∞and
␈↓ α←␈↓␈↓ β∂problem␈αsolving,␈αin␈αB. Meltzer␈αand␈αD. Michie␈α(eds.),␈α␈↓↓Machine␈αIntelligence
␈↓ α←␈↓↓␈↓ β∂6␈↓, Edinburgh University Press, Edinburgh, 1971, pp. 165-190.

␈↓"β␈↓ α←␈↓[Feldman72]
␈↓"β␈↓ α←␈↓␈↓ β∂Feldman,␈α∃J.,␈α⊗J. R.␈α∃Low,␈α⊗D. C.␈α∃Swinehart,␈α⊗and␈α∃R. H.␈α⊗Taylor:␈α∃Recent
␈↓ α←␈↓␈↓ β∂developments␈α∂in␈α∞␈↓¬SAIL␈↓,␈α∂an␈α∞␈↓¬ALGOL␈↓-based␈α∂language␈α∞for␈α∂artificial␈α∞intelligence,
␈↓ α←␈↓␈↓ β∂AIM 176, November 1972.

␈↓"β␈↓ α←␈↓[Fikes71]
␈↓"β␈↓ α←␈↓␈↓ β∂Fikes,␈αR. J.,␈αand␈αN. J.␈αNilsson:␈α
␈↓¬STRIPS␈↓--A␈αnew␈αapproach␈αto␈αthe␈α
application
␈↓ α←␈↓␈↓ β∂of␈α∪theorem␈α∀proving␈α∪to␈α∪problem␈α∀solving,␈α∪␈↓↓Artificial␈α∀Intelligence␈↓,␈α∪vol. 2,
␈↓ α←␈↓␈↓ β∂Winter 1971, pp. 189-208.

␈↓"β␈↓ α←␈↓[Fikes72]
␈↓"β␈↓ α←␈↓␈↓ β∂Fikes␈α⊗R. J.,␈α⊗P. E.␈α⊗Hart,␈α⊗and␈α⊗N. J.␈α⊗Nilsson:␈α⊗Learning␈α↔and␈α⊗executing
␈↓ α←␈↓␈↓ β∂generalized␈α⊂robot␈α⊃plans,␈α⊂␈↓↓Artificial␈α⊂Intelligence␈↓,␈α⊃vol. 3,␈α⊂Winter␈α⊃1972,␈α⊂pp.
␈↓ α←␈↓␈↓ β∂251-288.

␈↓"β␈↓ α←␈↓[Finkel74]
␈↓"β␈↓ α←␈↓␈↓ β∂Finkel,␈α↔R.,␈α_R.␈α↔Taylor,␈α↔R.␈α_Bolles,␈α↔R.␈α↔Paul,␈α_and␈α↔J.␈α↔Feldman:␈α_␈↓¬AL␈↓,␈α↔a
␈↓ α←␈↓␈↓ β∂programming system for automation, AIM 243, June 1975.
␈↓ α←␈↓␈↓REFERENCES␈↓ 
"285␈↓

␈↓"β␈↓ α←␈↓[Gelernter59]
␈↓"β␈↓ α←␈↓␈↓ β∂Gelernter,␈α∂H.:␈α∂Realization␈α⊂of␈α∂a␈α∂geometery-theorem␈α∂proving␈α⊂machine,␈α∂in
␈↓ α←␈↓␈↓ β∂E. A.␈α→Feigenbaum␈α→and␈α~J. Feldman␈α→(eds.),␈α→␈↓↓Computers␈α~and␈α→Thought␈↓,
␈↓ α←␈↓␈↓ β∂McGraw-Hill, New York, 1963, pp. 134-152.

␈↓"β␈↓ α←␈↓[Goldstein74]
␈↓"β␈↓ α←␈↓␈↓ β∂Goldstein,␈α
I.:␈α
Understanding␈αsimple␈α
picture␈α
programs,␈α
AI-TR-294,␈αMIT,
␈↓ α←␈↓␈↓ β∂September 1974.

␈↓"β␈↓ α←␈↓[Green69]
␈↓"β␈↓ α←␈↓␈↓ β∂Green,␈αC. C.:␈α
The␈αapplication␈α
of␈αtheorem␈α
proving␈αto␈α
question␈αanswering
␈↓ α←␈↓␈↓ β∂systems, AIM 96, August 1969.

␈↓"β␈↓ α←␈↓[Green74]
␈↓"β␈↓ α←␈↓␈↓ β∂Green,␈αC. C.,␈αR. J.␈αWaldinger,␈α
D. R.␈αBarstow,␈αR.␈αElschlager,␈α
D. B.␈αLenat,
␈↓ α←␈↓␈↓ β∂B. P.␈α
McCune,␈α
D. E.␈αShaw,␈α
L. I.␈α
Steinberg:␈αProgress␈α
report␈α
on␈αprogram-
␈↓ α←␈↓␈↓ β∂understanding systems, AIM 240, August 1974.

␈↓"β␈↓ α←␈↓[Gregory66]
␈↓"β␈↓ α←␈↓␈↓ β∂Gregory, R. L.:  ␈↓↓Eye and Brain␈↓, McGraw-Hill, New York, 1966.

␈↓"β␈↓ α←␈↓[Guzman68]
␈↓"β␈↓ α←␈↓␈↓ β∂Guzman,␈α∩A.:␈α∩Computer␈α∩recognition␈α∩of␈α∩3-D␈α∩objects␈α∩in␈α∩a␈α∩visual␈α∩scene,
␈↓ α←␈↓␈↓ β∂MAC-TR-59, MIT, December 1968.

␈↓"β␈↓ α←␈↓[Hansen74]
␈↓"β␈↓ α←␈↓␈↓ β∂Hansen,␈α∀G.:␈α∀Adaptive␈α∀systems␈α∪for␈α∀dynamic␈α∀run-time␈α∀optimization␈α∪of
␈↓ α←␈↓␈↓ β∂programs, doctoral dissertation, CMU, March 1974.

␈↓"β␈↓ α←␈↓[Harre70]
␈↓"β␈↓ α←␈↓␈↓ β∂Harre,␈α∞R.:␈α∞␈↓↓The␈α∞Principles␈α∞of␈α∞Scientific␈α∞Thinking␈↓,␈α∞University␈α∂of␈α∞Chicago
␈↓ α←␈↓␈↓ β∂Press, Chicago, 1970.

␈↓"β␈↓ α←␈↓[Hart75]
␈↓"β␈↓ α←␈↓␈↓ β∂Hart,␈α
P. E.:␈α
Progress␈α
on␈α
a␈α
computer-based␈α
consultant,␈α
␈↓↓4IJCAI␈↓,␈α
1975,␈αpp.
␈↓ α←␈↓␈↓ β∂831-841.

␈↓"β␈↓ α←␈↓[Hayes-Roth76]
␈↓"β␈↓ α←␈↓␈↓ β∂Hayes-Roth,␈α F.,␈α and␈α∨J. McDermott:␈α Knowledge␈α acquisition␈α∨from
␈↓ α←␈↓␈↓ β∂structural␈αdescriptions,␈αTech.␈αRep.,␈αComputer␈αScience␈αDepartment,␈αCMU,
␈↓ α←␈↓␈↓ β∂February 1976.

␈↓"β␈↓ α←␈↓[Hewitt71]
␈↓"β␈↓ α←␈↓␈↓ β∂Hewitt,␈αC.:␈αProcedural␈αsemantics--models␈αof␈αprocedures␈αand␈αthe␈αteaching
␈↓ α←␈↓␈↓ β∂of␈αprocedures,␈α␈↓↓Natural␈αLanguage␈αProcessing␈↓␈α(Courant␈αComputer␈αScience
␈↓ α←␈↓␈↓ β∂Symposium), vol. 8, 1971, pp. 331-350.
␈↓ α←␈↓␈↓286␈↓ 	0REFERENCES␈↓

␈↓"β␈↓ α←␈↓[Hewitt72]
␈↓"β␈↓ α←␈↓␈↓ β∂Hewitt,␈α↔C.:␈α↔Description␈α_and␈α↔theoretical␈α↔analysis␈α↔of␈α_␈↓¬PLANNER␈↓,␈α↔doctoral
␈↓ α←␈↓␈↓ β∂dissertation, Department of Mathematics, MIT, 1972.

␈↓"β␈↓ α←␈↓[Hewitt75]
␈↓"β␈↓ α←␈↓␈↓ β∂Hewitt,␈α∪C.,␈α∀and␈α∪B. Smith:␈α∪Toward␈α∀a␈α∪programmer's␈α∀apprentice,␈α∪␈↓↓IEEE
␈↓ α←␈↓↓␈↓ β∂Transactions on Software Engineering␈↓, SE-1, March 1975, pp. 26-45.

␈↓"β␈↓ α←␈↓[Howe73]
␈↓"β␈↓ α←␈↓␈↓ β∂Howe,␈α∩W. J.: ␈α∩Computer-assisted␈α∩design␈α∩of␈α∩complex␈α∩organic␈α⊃syntheses,
␈↓ α←␈↓␈↓ β∂doctoral␈α
dissertation,␈α
Harvard␈α
University,␈α
Cambridge,␈α
1973␈α
(␈↓↓Dissertation
␈↓ α←␈↓↓␈↓ β∂Abstracts␈↓, P5207 B, 33/11).

␈↓"β␈↓ α←␈↓[Interaction72]
␈↓"β␈↓ α←␈↓␈↓ β∂Interaction␈α∞Associates:␈α
␈↓↓Strategy␈α∞Notebook␈↓,␈α
Interaction␈α∞Asociates,␈α∞Inc.,␈α
San
␈↓ α←␈↓␈↓ β∂Francisco, Calif., 1972.

␈↓"β␈↓ α←␈↓[Johnson75]
␈↓"β␈↓ α←␈↓␈↓ β∂Johnson,␈α∂H.␈α∂R.:␈α∂A␈α∂schema␈α∂report␈α∂facility␈α∂for␈α∂a␈α∂CODASYL-based␈α∞data
␈↓ α←␈↓␈↓ β∂definition␈α∞language,␈α
in␈α∞B. C. M.␈α
Dougue␈α∞and␈α
G. M.␈α∞Nijssen␈α∞(eds.),␈α
␈↓↓Data
␈↓ α←␈↓↓␈↓ β∂Base Description␈↓, American Elsevier, New York, 1975, pp. 299-328.

␈↓"β␈↓ α←␈↓[Kulikowski73]
␈↓"β␈↓ α←␈↓␈↓ β∂Kulikowski,␈α↔C. A.,␈α↔S. Weiss,␈α↔and␈α↔A. Saifr:␈α↔Glaucoma␈α_diagnosis␈α↔and
␈↓ α←␈↓␈↓ β∂therapy␈α⊗by␈α⊗computer,␈α⊗␈↓↓Proceedings␈α⊗of␈α⊗Annual␈α⊗Meeting␈α⊗of␈α⊗Assoc.␈α∃for
␈↓ α←␈↓↓␈↓ β∂Research in Vision and Opthamology␈↓, May 1973.

␈↓"β␈↓ α←␈↓[Learning76]
␈↓"β␈↓ α←␈↓␈↓ β∂Learning␈αResearch␈αGroup:␈α
Personal␈αdynamic␈αmedia,␈αXerox␈α
PARC,␈αPalo
␈↓ α←␈↓␈↓ β∂Alto, Calif., 1976.

␈↓"β␈↓ α←␈↓[Lesser74]
␈↓"β␈↓ α←␈↓␈↓ β∂Lesser,␈αV.␈α
R.,␈αR. D.␈αFennell,␈α
L. D.␈αErman,␈α
and␈αD. R.␈αReddy:␈α
Organization
␈↓ α←␈↓␈↓ β∂of␈α∩the␈α∩␈↓¬HEARSAY␈α∩II␈↓␈α∩speech␈α∩understanding␈α∩system,␈α∩␈↓↓IEEE␈α∩Transactions␈α⊃on
␈↓ α←␈↓↓␈↓ β∂Acoustics,␈α∂Speech,␈α∂and␈α⊂Signal␈α∂Processing␈↓,␈α∂ASSP-23,␈α∂February␈α⊂1975,␈α∂pp.
␈↓ α←␈↓␈↓ β∂11-23.

␈↓"β␈↓ α←␈↓[Liskov74]
␈↓"β␈↓ α←␈↓␈↓ β∂Liskov,␈α→B.,␈α→and␈α→S.␈α→Zilles:␈α→Programming␈α→with␈α→abstract␈α~data␈α→types,
␈↓ α←␈↓␈↓ β∂␈↓↓SIGPLAN Notices␈↓, April 1974.

␈↓"β␈↓ α←␈↓[Low74]
␈↓"β␈↓ α←␈↓␈↓ β∂Low␈α
J.:␈α∞Automatic␈α
coding: ␈α∞Choice␈α
of␈α
data␈α∞structures,␈α
AIM␈α∞242,␈α
August
␈↓ α←␈↓␈↓ β∂1974.
␈↓ α←␈↓␈↓REFERENCES␈↓ 
"287␈↓

␈↓"β␈↓ α←␈↓[Lukasiewisz70]
␈↓"β␈↓ α←␈↓␈↓ β∂Lukasiewicz,␈αJ.:␈αA␈αnumerical␈αinterpretation␈αof␈αthe␈αtheory␈αof␈αpropositions,
␈↓ α←␈↓␈↓ β∂in L. Borkowski (ed.), ␈↓↓Jan Lukasiewicz:  Selected Works␈↓, 1970.

␈↓"β␈↓ α←␈↓[MACSYMA74]
␈↓"β␈↓ α←␈↓␈↓ β∂The␈α␈↓¬MACSYMA␈↓␈αreference␈αmanual,␈αThe␈αMATHLAB␈αGroup,␈αMIT,␈αSeptember
␈↓ α←␈↓␈↓ β∂1974.

␈↓"β␈↓ α←␈↓[Manna69]
␈↓"β␈↓ α←␈↓␈↓ β∂Manna,␈α↔Z.:␈α↔Correctness␈α↔of␈α↔programs,␈α↔␈↓↓Journal␈α↔of␈α↔Computer␈α⊗Systems
␈↓ α←␈↓↓␈↓ β∂Sciences␈↓, May 1969.

␈↓"β␈↓ α←␈↓[McDermott74]
␈↓"β␈↓ α←␈↓␈↓ β∂McDermott,␈α∃D.:␈α∃Assimilation␈α∃of␈α∃new␈α∃information,␈α⊗AI-TR-291,␈α∃MIT,
␈↓ α←␈↓␈↓ β∂February 1974.

␈↓"β␈↓ α←␈↓[McLeod76]
␈↓"β␈↓ α←␈↓␈↓ β∂McLeod,␈α∂D.␈α⊂J.:␈α∂High␈α∂level␈α⊂domain␈α∂definition␈α∂in␈α⊂a␈α∂relational␈α⊂data␈α∂base
␈↓ α←␈↓␈↓ β∂system, ␈↓↓SIGPLAN Notices␈↓, no. 1, April 1976, pp. 47-57.

␈↓"β␈↓ α←␈↓[Miller75]
␈↓"β␈↓ α←␈↓␈↓ β∂Miller,␈α
P. B.:␈α
Strategy␈α
selection␈α∞in␈α
medical␈α
diagnosis,␈α
Project␈α∞MAC␈α
TR-
␈↓ α←␈↓␈↓ β∂153, MIT, September 1975.

␈↓"β␈↓ α←␈↓[Minsky68]
␈↓"β␈↓ α←␈↓␈↓ β∂Minsky,␈α∪M.␈α∪(ed.):␈α∪␈↓↓Semantic␈α∪Information␈α∪Processing␈↓,␈α∪MIT␈α∀Press,␈α∪MIT,
␈↓ α←␈↓␈↓ β∂1968.

␈↓"β␈↓ α←␈↓[Minsky74]
␈↓"β␈↓ α←␈↓␈↓ β∂Minsky,␈α∂M.:␈α⊂A␈α∂framework␈α∂for␈α⊂representing␈α∂knowledge,␈α∂MIT␈α⊂AI␈α∂Memo
␈↓ α←␈↓␈↓ β∂306, June 1974.

␈↓"β␈↓ α←␈↓[Mitchell70]
␈↓"β␈↓ α←␈↓␈↓ β∂Mitchell,␈α∩J.␈α⊃G.:␈α∩The␈α∩design␈α⊃and␈α∩construction␈α⊃of␈α∩flexible␈α∩and␈α⊃efficient
␈↓ α←␈↓␈↓ β∂interactive␈α∀programming␈α∪systems,␈α∀doctoral␈α∪dissertation,␈α∀Department␈α∪of
␈↓ α←␈↓␈↓ β∂Computer Science, CMU, June 1970.

␈↓"β␈↓ α←␈↓[Newell59]
␈↓"β␈↓ α←␈↓␈↓ β∂Newell,␈α
A.,␈αJ. C.␈α
Shaw,␈αand␈α
H. A.␈αSimon:␈α
A␈αvariety␈α
of␈αintelligent␈α
learning
␈↓ α←␈↓␈↓ β∂in␈α∞a␈α
general␈α∞problem-solver,␈α
in␈α∞M. Yovitts␈α
and␈α∞S.␈α
Cameron␈α∞(eds.),␈α
␈↓↓Self-
␈↓ α←␈↓↓␈↓ β∂organizing Systems␈↓, Pergamon, New York, pp. 153-189.

␈↓"β␈↓ α←␈↓[Newell61]
␈↓"β␈↓ α←␈↓␈↓ β∂Newell,␈α∀A.,␈α∀and␈α∪H. A.␈α∀Simon:␈α∀␈↓¬GPS␈↓,␈α∪a␈α∀program␈α∀that␈α∀simulates␈α∪human
␈↓ α←␈↓␈↓ β∂thought,␈α∪in␈α∪E. A.␈α∩Feigenbaum␈α∪and␈α∪J.␈α∩Feldman␈α∪(eds.),␈α∪␈↓↓Computers␈α∩and
␈↓ α←␈↓↓␈↓ β∂Thought␈↓, McGraw-Hill, New York, pp. 279-296.
␈↓ α←␈↓␈↓288␈↓ 	0REFERENCES␈↓

␈↓"β␈↓ α←␈↓[Newell69]
␈↓"β␈↓ α←␈↓␈↓ β∂Newell,␈αA.:␈αHeuristic␈αprogramming;␈αill-structured␈αproblems,␈αin␈αAronofsky
␈↓ α←␈↓␈↓ β∂(ed.), ␈↓↓Progress in Operations Research␈↓, vol. 3, 1969, pp.  362-414.

␈↓"β␈↓ α←␈↓[Newell72]
␈↓"β␈↓ α←␈↓␈↓ β∂Newell,␈α↔A.,␈α↔and␈α↔H.␈α⊗Simon:␈α↔␈↓↓Human␈α↔Problem␈α↔Solving␈↓,␈α⊗Prentice-Hall,
␈↓ α←␈↓␈↓ β∂Englewood Cliffs, New Jersey, 1972.

␈↓"β␈↓ α←␈↓[Norman75]
␈↓"β␈↓ α←␈↓␈↓ β∂Norman,␈α∞D.␈α∞A.,␈α
and␈α∞D.␈α∞E.␈α∞Rumelhart:␈α
␈↓↓Explorations␈α∞in␈α∞Cognition␈↓,␈α∞W.␈α
H.
␈↓ α←␈↓␈↓ β∂Freeman, San Francisco, 1975.

␈↓"β␈↓ α←␈↓[Parnas72]
␈↓"β␈↓ α←␈↓␈↓ β∂Parnas,␈α∂D.␈α∂L.:␈α⊂On␈α∂the␈α∂criteria␈α∂to␈α⊂be␈α∂used␈α∂in␈α∂decomposing␈α⊂systems␈α∂into
␈↓ α←␈↓␈↓ β∂modules, ␈↓↓CACM␈↓, vol. 15, December 1972, pp. 1053-1058.

␈↓"β␈↓ α←␈↓[Parnas75]
␈↓"β␈↓ α←␈↓␈↓ β∂Parnas,␈αD.␈αL.,␈αand␈αD. P.␈αSiewiorek:␈αUse␈αof␈αthe␈αconcept␈αof␈αtransparency␈αin
␈↓ α←␈↓␈↓ β∂the␈αdesign␈αof␈αhierarchically␈αstructured␈αsystems,␈α␈↓↓CACM␈↓,␈αvol.␈α18,␈αJuly␈α1975,
␈↓ α←␈↓␈↓ β∂pp. 401-408.

␈↓"β␈↓ α←␈↓[Polya54]
␈↓"β␈↓ α←␈↓␈↓ β∂Polya, G.: ␈↓↓How to Solve it␈↓, McGraw-Hill, Princeton, New Jersey, 1954.

␈↓"β␈↓ α←␈↓[Pople75]
␈↓"β␈↓ α←␈↓␈↓ β∂Pople,␈αH.,␈αJ.␈αMeyers,␈αand␈αR.␈α
Miller:␈α␈↓¬DIALOG␈↓,␈αa␈αmodel␈αof␈αdiagnostic␈αlogic␈α
for
␈↓ α←␈↓␈↓ β∂internal␈α⊂medicine,␈α⊃␈↓↓4IJCAI␈↓,␈α⊂1975,␈α⊃pp.␈α⊂ 848-855.␈α⊃ (The␈α⊂system␈α⊃has␈α⊂since
␈↓ α←␈↓␈↓ β∂been renamed ␈↓¬INTERNIST␈↓.)

␈↓"β␈↓ α←␈↓[Post43]
␈↓"β␈↓ α←␈↓␈↓ β∂Post,␈α_E.:␈α→Formal␈α_reductions␈α_of␈α→the␈α_general␈α→combinatorial␈α_problem,
␈↓ α←␈↓␈↓ β∂␈↓↓American Journal of Math␈↓, vol. 65, 1943, pp. 197-268.
␈↓"β␈↓ α←␈↓␈↓ β∂For␈α⊂an␈α∂introduction␈α⊂to␈α∂the␈α⊂general␈α∂principles␈α⊂involved,␈α⊂see␈α∂M. Minksy,
␈↓ α←␈↓␈↓ β∂␈↓↓Computation: ␈α∩Finite␈α∩and␈α∩infinite␈α∩machines␈↓,␈α∩Prentice␈α∩Hall,␈α∩Englewwod
␈↓ α←␈↓␈↓ β∂Cliffs, New Jersey, 1967, chap. 12.

␈↓"β␈↓ α←␈↓[Reddy73]
␈↓"β␈↓ α←␈↓␈↓ β∂Reddy,␈α∞R.:␈α∞The␈α∞␈↓¬HEARSAY␈↓␈α∞speech␈α∞understanding␈α∞system,␈α∞␈↓↓3IJCAI␈↓,␈α∂1973,␈α∞pp.
␈↓ α←␈↓␈↓ β∂185-199.

␈↓"β␈↓ α←␈↓[Reiger74]
␈↓"β␈↓ α←␈↓␈↓ β∂Reiger,␈α∩C.␈α∩J.:␈α∩Conceptual␈α∩memory: ␈α∩A␈α∩theory␈α∩and␈α∪computer␈α∩program,
␈↓ α←␈↓␈↓ β∂AIM 233, July 1974.
␈↓ α←␈↓␈↓REFERENCES␈↓ 
"289␈↓

␈↓"β␈↓ α←␈↓[Roberts63]
␈↓"β␈↓ α←␈↓␈↓ β∂Roberts,␈α
L.␈α
G.:␈α
Machine␈α
perception␈αof␈α
3-D␈α
solids,␈α
Tech.␈α
Rep.␈α315,␈α
Lincoln
␈↓ α←␈↓␈↓ β∂Labs, MIT, May 1963.

␈↓"β␈↓ α←␈↓[Rubin75]
␈↓"β␈↓ α←␈↓␈↓ β∂Rubin,␈αA.␈αD.:␈αHypothesis␈αformation␈αand␈αevaluation␈αin␈αmedical␈αdiagnosis,
␈↓ α←␈↓␈↓ β∂AI-TR-316, MIT, January 1975.

␈↓"β␈↓ α←␈↓[Rumelhart73]
␈↓"β␈↓ α←␈↓␈↓ β∂Rumelhart,␈α∩D.␈α⊃E.,␈α∩and␈α⊃D. A.␈α∩Norman:␈α⊃Active␈α∩semantic␈α⊃networks␈α∩as␈α⊃a
␈↓ α←␈↓␈↓ β∂model of human memory, ␈↓↓3IJCAI␈↓, 1973, pp. 450-458.

␈↓"β␈↓ α←␈↓[Rulifson72]
␈↓"β␈↓ α←␈↓␈↓ β∂Rulifson,␈α⊃J.␈α⊂F.,␈α⊃J. A.␈α⊃Derksen,␈α⊂and␈α⊃R. J.␈α⊂Waldinger:␈α⊃␈↓¬QA4␈↓:␈α⊃A␈α⊂procedural
␈↓ α←␈↓␈↓ β∂calculus␈α∞for␈α∞intuitive␈α
reasoning,␈α∞Tech.␈α∞ Note␈α
73,␈α∞SRI␈α∞International,␈α
Palo
␈↓ α←␈↓␈↓ β∂Alto, Calif., November 1972.

␈↓"β␈↓ α←␈↓[Rumelhart73]
␈↓"β␈↓ α←␈↓␈↓ β∂Rumelhart,␈α∩D. E.,␈α∪and␈α∩D. A.␈α∩Norman:␈α∪Active␈α∩semantic␈α∩networks␈α∪as␈α∩a
␈↓ α←␈↓␈↓ β∂model of human memory, ␈↓↓3IJCAI␈↓, 1973, pp. 412-422.

␈↓"β␈↓ α←␈↓[Sacerdoti73]
␈↓"β␈↓ α←␈↓␈↓ β∂Sacerdoti,␈α∩E.:␈α∩Planning␈α∩in␈α∩a␈α∩hierarchy␈α∩of␈α∩abstraction␈α∩spaces,␈α⊃␈↓↓3IJCAI␈↓,
␈↓ α←␈↓␈↓ β∂1973, pp. 412-422.

␈↓"β␈↓ α←␈↓[Sacerdoti77]
␈↓"β␈↓ α←␈↓␈↓ β∂Sacerdoti,␈α⊃E.:␈α⊃␈↓↓A␈α⊃Structure␈α⊃for␈α⊃Plans␈α⊃and␈α⊃Behavior␈↓,␈α∩American␈α⊃Elsevier,
␈↓ α←␈↓␈↓ β∂New York, in press.

␈↓"β␈↓ α←␈↓[Samet75]
␈↓"β␈↓ α←␈↓␈↓ β∂Samet,␈α
H.:␈αAutomatically␈α
proving␈α
the␈αcorrectness␈α
of␈αtranslations␈α
involving
␈↓ α←␈↓␈↓ β∂optimized code, AIM 259, May 1975.

␈↓"β␈↓ α←␈↓[Samuel67]
␈↓"β␈↓ α←␈↓␈↓ β∂Samuel,␈α∩A.␈α∩L.:␈α∩Some␈α∩studies␈α∪in␈α∩machine␈α∩learning␈α∩using␈α∩the␈α∪game␈α∩of
␈↓ α←␈↓␈↓ β∂checkers␈αII--recent␈αprogress,␈α
␈↓↓IBM␈αJournal␈αof␈αResearch␈α
and␈αDevelopment␈↓,
␈↓ α←␈↓␈↓ β∂vol. 11, 1967, pp. 601-617.

␈↓"β␈↓ α←␈↓[Sandewall75]
␈↓"β␈↓ α←␈↓␈↓ β∂Sandewall,␈α
E.:␈α
Ideas␈α
about␈α
managment␈αof␈α
␈↓¬LISP␈↓␈α
data␈α
bases,␈α
␈↓↓4IJCAI␈↓,␈α1975,
␈↓ α←␈↓␈↓ β∂pp. 585-592.

␈↓"β␈↓ α←␈↓[Shaw75]
␈↓"β␈↓ α←␈↓␈↓ β∂Shaw,␈α∩D.,␈α∩W.␈α∩Swartout,␈α∩and␈α∩C.␈α∩Green:␈α∩Inferring␈α∩␈↓¬LISP␈↓␈α∩programs␈α∩from
␈↓ α←␈↓␈↓ β∂examples, ␈↓↓4IJCAI␈↓, 1975, pp. 260-267.
␈↓ α←␈↓␈↓290␈↓ 	0REFERENCES␈↓

␈↓"β␈↓ α←␈↓[Shortliffe75a]
␈↓"β␈↓ α←␈↓␈↓ β∂Shortliffe,␈α∞E.␈α∂H.,␈α∞R.␈α∂Davis,␈α∞B. G.␈α∞Buchanan,␈α∂S. G.␈α∞Axline,␈α∂C. C.␈α∞Green,
␈↓ α←␈↓␈↓ β∂and␈αS. N.␈αCohen:␈αComputer-based␈αconsultations␈αin␈αclinical␈αtherapeutics--
␈↓ α←␈↓␈↓ β∂explanation␈α~and␈α≠rule␈α~acquisition␈α~capabilities␈α≠of␈α~the␈α≠␈↓¬MYCIN␈↓␈α~system,
␈↓ α←␈↓␈↓ β∂␈↓↓Computers and Biomedical Research␈↓, vol. 8, 1975, pp. 303-320.

␈↓"β␈↓ α←␈↓[Shortliffe75b]
␈↓"β␈↓ α←␈↓␈↓ β∂Shortliffe,␈α
E.␈α
H.,␈α
and␈α
B. G.␈α
Buchanan:␈α
A␈α
model␈α
of␈α
inexact␈α
reasoning␈α
in
␈↓ α←␈↓␈↓ β∂medicine, ␈↓↓Mathematical Biosciences␈↓, vol. 23, 1975, pp. 351-379.

␈↓"β␈↓ α←␈↓[Shortliffe76]
␈↓"β␈↓ α←␈↓␈↓ β∂Shortliffe,␈α≠E.␈α≠H.:␈α≠␈↓↓MYCIN: ␈α≠Computer-based␈α≠Medical␈α≠Consultations␈↓,
␈↓ α←␈↓␈↓ β∂American Elsevier, New York, 1976.

␈↓"β␈↓ α←␈↓[Simon73]
␈↓"β␈↓ α←␈↓␈↓ β∂Simon,␈αH.:␈α
The␈αstructure␈α
of␈αill-structured␈α
problems,␈α␈↓↓Artificial␈α
Intelligence␈↓,
␈↓ α←␈↓␈↓ β∂vol. 4, 1973, pp.181-201.

␈↓"β␈↓ α←␈↓[Sophocles27]
␈↓"β␈↓ α←␈↓␈↓ β∂Sophocles,␈α∞Oedipus␈α
the␈α∞King␈α
(427␈α∞BC),␈α
in␈α∞Greene␈α
and␈α∞Lattimore␈α
(eds.),
␈↓ α←␈↓␈↓ β∂␈↓↓Greek Tragedies␈↓, vol. 1, University of Chicago Press, Chicago, 1960.

␈↓"β␈↓ α←␈↓[Spitzen75]
␈↓"β␈↓ α←␈↓␈↓ β∂Spitzen,␈α∪J.,␈α∀and␈α∪B. Wegbreit:␈α∪The␈α∀verification␈α∪and␈α∪synthesis␈α∀of␈α∪data
␈↓ α←␈↓␈↓ β∂structures, ␈↓↓Acta Informatica␈↓, vol. 4, 1975, pp. 127-144.

␈↓"β␈↓ α←␈↓[Stonebreaker75]
␈↓"β␈↓ α←␈↓␈↓ β∂Stonebreaker,␈α∂M.:␈α∂Implementation␈α⊂of␈α∂integrity␈α∂constraints␈α∂and␈α⊂views␈α∂by
␈↓ α←␈↓␈↓ β∂modification, ␈↓↓Proc. SIGMOD Conf.␈↓, 1975, pp. 65-78.

␈↓"β␈↓ α←␈↓[Sussman75]
␈↓"β␈↓ α←␈↓␈↓ β∂Sussman,␈α∃G.:␈α∃␈↓↓A␈α∃Computational␈α∀Model␈α∃of␈α∃Skill␈α∃Acquisition,␈↓␈α∀American
␈↓ α←␈↓␈↓ β∂Elsevier, New York, 1975.

␈↓"β␈↓ α←␈↓[Suzuki76]
␈↓"β␈↓ α←␈↓␈↓ β∂Suzuki,␈α↔N.:␈α⊗Automatic␈α↔verification␈α↔of␈α⊗programs␈α↔with␈α↔complex␈α⊗data
␈↓ α←␈↓␈↓ β∂structures, AIM 279, February 1976.

␈↓"β␈↓ α←␈↓[Tversky74]
␈↓"β␈↓ α←␈↓␈↓ β∂Tversky,␈αA.,␈αand␈αD. Kahneman:␈αJudgment␈αunder␈αuncertainty: ␈αHeuristics
␈↓ α←␈↓␈↓ β∂and biases, ␈↓↓Science␈↓, vol. 185, September 18 1974, pp. 1129-1131.

␈↓"β␈↓ α←␈↓[van Melle74]
␈↓"β␈↓ α←␈↓␈↓ β∂van␈α⊂Melle,␈α⊃W.:␈α⊂Would␈α⊂you␈α⊃like␈α⊂advice␈α⊂on␈α⊃another␈α⊂horn,␈α⊃␈↓¬MYCIN␈↓␈α⊂project
␈↓ α←␈↓␈↓ β∂internal␈α∞working␈α∞paper,␈α∞Stanford␈α∞University,␈α∞Stanford,␈α∞Calif.,␈α∞December
␈↓ α←␈↓␈↓ β∂1974.
␈↓ α←␈↓␈↓REFERENCES␈↓ 
"291␈↓

␈↓"β␈↓ α←␈↓[Waldinger74]
␈↓"β␈↓ α←␈↓␈↓ β∂Waldinger,␈α⊂R.,␈α⊂and␈α⊃K. N.␈α⊂Levitt:␈α⊂Reasoning␈α⊂about␈α⊃programs,␈α⊂␈↓↓Artificial
␈↓ α←␈↓↓␈↓ β∂Intelligence␈↓, vol. 5, Fall 1974, pp. 235-316.

␈↓"β␈↓ α←␈↓[Waltz72]
␈↓"β␈↓ α←␈↓␈↓ β∂Waltz,␈αD.:␈αGenerating␈αsemantic␈αdescriptions␈αfrom␈αdrawings␈αof␈αscenes␈α
with
␈↓ α←␈↓␈↓ β∂shadows, AI-TR-271, MIT, November 1972.

␈↓"β␈↓ α←␈↓[Warshall62]
␈↓"β␈↓ α←␈↓␈↓ β∂Warshall,␈α⊃S.:␈α⊃A␈α⊃theorem␈α⊃on␈α⊃Boolean␈α⊃matrices,␈α⊃␈↓↓JACM␈↓,␈α⊃vol. 9,␈α⊃January
␈↓ α←␈↓␈↓ β∂1962, pp. 11-12.

␈↓"β␈↓ α←␈↓[Waterman70]
␈↓"β␈↓ α←␈↓␈↓ β∂Waterman,␈αD.␈α
A.:␈αGeneralization␈αlearning␈α
techniques␈αfor␈α
automating␈αthe
␈↓ α←␈↓␈↓ β∂learning of heuristics, ␈↓↓Artificial Intelligence␈↓, vol. 1, 1970, pp. 121-170.

␈↓"β␈↓ α←␈↓[Waterman77]
␈↓"β␈↓ α←␈↓␈↓ β∂Waterman,␈α≥D. A.:␈α≥Exemplary␈α≤programming,␈α≥in␈α≥D. Waterman␈α≤and
␈↓ α←␈↓␈↓ β∂R. Hayes-Roth␈α(eds.),␈α␈↓↓Pattern-directed␈α
Inference␈αSystems␈↓,␈αAcademic␈α
Press,
␈↓ α←␈↓␈↓ β∂in press.

␈↓"β␈↓ α←␈↓[Wickelgren74]
␈↓"β␈↓ α←␈↓␈↓ β∂Wickelgren,␈α_W.␈α↔A.:␈α_␈↓↓How␈α↔to␈α_Solve␈α↔Problems␈↓,␈α_W.␈α↔H.␈α_Freeman,␈α↔San
␈↓ α←␈↓␈↓ β∂Francisco, 1974.

␈↓"β␈↓ α←␈↓[Winograd72]
␈↓"β␈↓ α←␈↓␈↓ β∂Winograd,␈α
T.: ␈α
␈↓↓Understanding␈α
Natural␈α
Language␈↓,␈α
Academic␈α
Press,␈α
New
␈↓ α←␈↓␈↓ β∂York, 1972.

␈↓"β␈↓ α←␈↓[Winograd74]
␈↓"β␈↓ α←␈↓␈↓ β∂Winograd,␈α→T.:␈α~Breaking␈α→the␈α~complexity␈α→barrier,␈α~again,␈α→␈↓↓SIGPLAN
␈↓ α←␈↓↓␈↓ β∂Notices␈↓, no. 1, January 1974.

␈↓"β␈↓ α←␈↓[Winograd75]
␈↓"β␈↓ α←␈↓␈↓ β∂Winograd,␈α∃T.: ␈α∃Frame␈α∃representations␈α∃and␈α∃the␈α∃procedural/declarative
␈↓ α←␈↓␈↓ β∂controversy,␈α∂in␈α∞D. N.␈α∂Bobrow␈α∞and␈α∂A.␈α∞Collins␈α∂(eds.),␈α∂␈↓↓Representation␈α∞and
␈↓ α←␈↓↓␈↓ β∂Understanding␈↓, Academic Press, New York, 1975.

␈↓"β␈↓ α←␈↓[Winston70]
␈↓"β␈↓ α←␈↓␈↓ β∂Winston,␈α⊂P.␈α⊂H.:␈α⊂Learning␈α⊂structural␈α⊂descriptions␈α⊂from␈α⊃examples,␈α⊂MAC
␈↓ α←␈↓␈↓ β∂TR-76, MIT, September 1970.
␈↓ α←␈↓␈↓292␈↓ 	0REFERENCES␈↓

␈↓"β␈↓ α←␈↓␈↓ ε?``Paul Cohen agrees with me:
␈↓"β␈↓ α←␈↓␈↓ ε?Everything␈α_that's␈α_ever␈α_been␈α_done␈α↔is
␈↓ α←␈↓␈↓ ε?irrelevant.''
␈↓"β␈↓ α←␈↓␈↓ π7Doug Lenat
␈↓"β␈↓ α←␈↓␈↓ π711 March 1976