perm filename DECUS.TLK[RDG,DBL] blob
sn#696139 filedate 1983-01-13 generic text, type C, neo UTF8
COMMENT ⊗ VALID 00015 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00002 00002 This presentation will be 2 distinct,
C00003 00003 MYCIN SCENARIO
C00004 00004 Narrow down to those of KE approach
C00005 00005 DEC's activities
C00007 00006 In general, some proof procedure.
C00008 00007 * Modus Monens
C00009 00008 Extendability
C00010 00009 System can explain reasoning -
C00011 00010 SIMPLIFICATIONS
C00012 00011 MYCIN is interactive
C00013 00012 End of Part I.
C00014 00013 [[[ LEAVE DISPLAYED ]]]
C00015 00014 Automating KA
C00016 00015 1. Gross characteristics
C00017 ENDMK
C⊗;
This presentation will be 2 distinct,
but interconnected, talks.
Expert Systems --
What they are
Why worth studying
How they work (coarse)
When more?
Once central bottleneck -- Knowledge Acquisition -- solved.
(That is 2nd part)
-----
Meta Point -
Ask Questions!
Esp if too slow, or too fast
MYCIN SCENARIO
Begin talk w/story
[turn off projector]
General Practioner
examines patient,
sees possible meningitus
calls infectuous disease expert
Ensuing dialogue,
mixed initiative
unknowns - (spinal tap) test not run
uncertain - how high fever really is
GP can challenge --
why did you ask this?
how did you conclude that?
Diagnosis
- organisms responsible
Treatment
- to cover all remaining cases
- complete with caveats
----
Happily ever after
Narrow down to those of KE approach
(Report by Buchanan & Duda)
- judgmental (w/ formal justification)
- explain conclusions; justify results
- modify; incremental input
-------
Onto structural description
- internals needed for ↑
DEC's activities
As DEC users forum
As example of how big company ...
(IBM, TI, ... have similar story)
(From talk by S Polit, 26/Oct/82)
XCON in use
For misc DEC machines
Determines what else to be ordered
whether can be filled, w/in constraint
explains additional facts -- what is missed, why, ...
Gives assembly instructions
XSEL in use
- Interactive
- Its output to XCON
XSITE
- Room layout to physical layout
ISA
- when full system will be ready
-----
Sizer
- by Phil London
- to minimize energy
Layout Tool
- to fit components on chips, efficiently
-----
First Principles Reasoner
- w/ Randy Davis
- Structural descrption
Behaviour [Electrical vs Inference]
Principle of Violated Expectation
Explicit Fault Model
In general, some proof procedure.
Simple one: Backward Chaining
Here is simple case
(overly simplified --
we'll discuss deficiencies at end)
* Modus Monens
To prove B
Given A =>B,
Simply proof A
Notation - AND , OR
----
Easy to use these "chunks" for explanation
why/how conclusion reached
Extendability
* NOT simply decision tree
- no where explicit -- just drawn for pedagogy
* New branches
- added by simple FACTS RE: WORLD (rule)
- not hardwired;
no need to describe how it relates to world -- fits itself
* This rule applicable anyway you need to acquire $$.
System can explain reasoning -
how conclusion reached
Show subgoals achieved, and rules used
(This simple example only for pedagogic)
SIMPLIFICATIONS
SEARCH
Only One Direction of Search --
(Could be Forward, as wel as Backward)
[Dictated by topology of search spae - meta-level issues]
(Order of instantiation)
∃ many goals on "frontier";
Which goal to expand next?
If wrong, could lead BACKTRACK
or to ∞ descent.
Focus of Attention (esp when interactive)
-----
VARIABLES -- many things abstractable
Fred -> person
-----
Interaction
Focus of Att'n -
could give all facts re: x first -- or not
HUMAN ENGINEERING
Q/Aing, ...
MYCIN is interactive
User can ask why Mycin asked <x>.
or how conclusion reached
----
Easy to credit/blame a specific rule
- & fix system by changing rule, if appropriate!
(This tool solves large class of errors)
because RULES ARE MODULAR!
End of Part I.
Note complex rule?
Q: How did it get here?
A: KA! == topic of Part II
[[[ LEAVE DISPLAYED ]]]
Facts re: world -> expert -> KB
which KB uses to solve problem.
Why going so poorly?
This brings up most of problems w/ESs in general.
This is major research issue, @ S, etc.
"central bottleneck in ES construction"
Automating KA
- first by person as interface,
- gradually remove him from loop
1.1 Think in English, but talk in Lisp
Or Spatially, but type in sequentially
2.1 We know re: rules, blackboard models, predicate calculus
[Like Malinowski: If All have = hammer, all world looks like a nail.]
1. Gross characteristics
Classes
Relevant Properties
2. Fleshing out
[New instances]
3. Refinements
[Small twiddles, minor rearrangements]
----
E.g. Facts re: Organisms
Longevity
Host
Type - Viral, Bacterial, Fungal, ...