00100	.SEC EVALUATION AND IMPROVEMENTS
00200	
00300	.SS 7.1 Evaluation
00400		The primary aim in constructing this model  was  to  explore,
00500	develop,  test  and improve a theory having explanatory authenticity.
00600	To satisfy this aim the model must meet norms of internal consistency
00700	and  norms  of  external correspondence with observation. A secondary
00800	aim would involve pragmatic norms of application.  These aims are not
00900	unrelated but the first is more fundamental since useful applications
01000	require some degree of consistency and authenticity.
01100		As  emphasized  in  Chapter  2  a  model  in  the  form of an
01200	algorithm consists of a structure of functions  or  procedures  whose
01300	inner  workings  are  sufficient  to  reproduce  the outward symbolic
01400	behavior under consideration. The theory embodied  in  the  model  is
01500	revealed  in  the set of statements which illuminate  the connections
01600	betweeen input and output by  describing  how  the  structure  reacts
01700	under various circumstances.
01800		Theories have many  functions.  They  can  be  summarized  as
01900	follows [from Bunge?] .V
02000		(1)To systematize knowledge.
02100		(2)To  explain  facts  by  showing  how they are the entailed
02200		consequences of the systematizing hypotheses.
02300		(3)To increase knowledge by deriving new facts.
02400		(4)To enhance the testability of hypotheses by connecting
02500		them to observations.
02600		(5)To guide  research  by:                
02700			(a) posing fruitful problems
02800			(b) suggesting new data to gather
02900			(c) opening new lines of investigation
03000		(6)To map a portion of reality. .END
03100		It is a tall order for a  theory  to  fulfill  all  of  these
03200	functions.  In undeveloped fields we should be happy with even one of
03300	them.    Models  can  be  assigned  these  functions  when  they  are
03400	theoretical  in  type.   Our  model  was  intended primarily to serve
03500	functions (2) and (4), testable explanation.
03600		What  constitutes a satisfactory explanation has been treated
03700	in 2.1.  The "fit" or correspondence with phenomena as  indicated  by
03800	measurements and empirical tests indicate the degree of faithfuleness
03900	of the reproduction as described in Chapter 6.
04000		Decision procedures for a consensus acceptability of a  model
04100	sometimes  depend  not  so  much  on  truth,  an elusive state, as on
04200	whether a majority of the  relevant  expert  community  believes  the
04300	theory  or  model to approximate truth to some unknown and unknowable
04400	degree and be better than available nonoutlandish alternatives. Truth
04500	or  falsity cannot be proven with certainty but their presence can be
04600	assayed by some sort of critical assesment and deliberation.  We  can
04700	forgive models for being only nearly true. Validation is ultimately a
04800	private  experience.  A  theory  or  model  may  bring  cognitive  or
04900	pragmatic  comfort,  not because it is TRUE but because it represents
05000	an improvement over its contending rivals.  Cognitive  comfort  is  a
05100	type  of  intellectual  satisfaction  while pragmatic comfort accrues
05200	from applications to problems in order to make things  work  the  way
05300	humans  want  them  to  work  efficiently  in  practical  contexts of
05400	technological action. For the pragmatist a model is  a  means  to  an
05500	end;  for  the theoretician an explanatory model is an end in itself.
05600	It is hoped  that  this  paranoid  model  will  be  useful  in  least
05700	understanding  one  of  the  mysteries of human conduct, the paranoid
05800	mode and the paranoid streak which renders  so  many  susceptible  to
05900	idelogical delusions in which others appear as oppressors.
06000		It is commonly stated that an infinite number of theories  or
06100	models  can  account  for  the  same  data  of  observation.  Without
06200	questioning whether "infinite" means indefinitely large or just  more
06300	than  one,  we must allow for rival explanations. For a rival to be a
06400	live and tenable option it should  be  truly  alternative  (not  just
06500	saying  the same thing in a different way), have an appreciable prior
06600	probability, and be testable.
06700		Although  I  hold  that  faithful  reproduction,  fidelity as
06800	measured by indistinguishability, is the proper and  major  test  for
06900	the  adequacy  of  symbolic  models, it would be a bonus if our model
07000	could satisfy the function of  making possible  new  knowledge  through
07100	prediction. Like validation, the term "prediction" has many  meanings
07200	ranging  from  forecasts,  to  prognoses  to  prophecies  to  precise
07300	point-predictions such as an eclipse. To predict  is  to  announce  a
07400	fact  without  prior  knowledge of it. However one needs knowledge of
07500	the kind of fact expected, the conditions which produce  it  and  the
07600	circumstances   under   which  it  will  occur.  Accurate  long-range
07700	predictions characterize the ideal of celestial mechanics.  (But even
07800	astronomers,  with  their isolated and repetitive systems, have their
07900	troubles; Halley's comet in the 18th century arrived four days  later
08000	than  predicted  while in the 20th century it was eight days later!).
08100	Long-range predictions of individual  human  behavior  are  difficult
08200	because  (1)  sufficient  knowledge of initial conditions may require
08300	that we know the whole past history of a person-  something  not  yet
08400	achieved, (2) individuals do not remain
08500	isolated over the time stretch of the prediction  and  interact  with
08600	other  individuals of an unknown nature (3) life is a fortuitous flux
08700	of chance intersections of independent causal chains.  In  one  sense
08800	our  paranoid  model  makes moment-to-moment predictions and predicts
08900	true counterfactuals  about  behavior  in  a  psychiatric  interview.
09000	That is, if an interviewer  says  X  under  conditions  Y,  then  the
09100	model's  response  will  be  characterized by z1...zn. Counterfactual
09200	prediction  means  that  on  the  basis  of  observed behavior we are
09300	willing to assume the presence of unobserved behavior  potentials  in
09400	the  model's  repertoire  of  capabilities.  Predicting  new kinds of
09500	events or properties, instead of kinds we are already familiar  with,
09600	would  represent  a  surplus bonus. It would give both clinicians and
09700	investigators something to look for.
09800		This novelty could arise in  two  ways.    First,  the  model
09900	might demonstrate a property of the paranoid mode hitherto unobserved
10000	clinically.    In principle this could come  about  because  the  I/O
10100	behavior  of  the  model  is  a  consequence  of  a  large  number of
10200	interacting hypotheses and assumptions chosen  initially  to  explain
10300	frequently observed phenomena.    When the elements of such a complex
10400	conjunction  interact  with  highly  variable  inputs  they  generate
10500	consequences  in  addition  to  those  they were designed to explain.
10600	Whether any of these consequences are significant  or  characteristic
10700	of the paranoid mode remains a subject for future study.
10800		It is also possible that a new property of  paranoia  may  be
10900	discovered  in  the  clinical  interview, although perhaps everything
11000	that can be said about it has been  said.  If  a  new  property  were
11100	found, a search for it might be conducted in the model's behavior. if
11200	successful, this would add to the model' acceptability.
11300		A  second novelty might arise in the behavior of the model in
11400	some new situation.   Since it is designed to simulate  communicative
11500	behavior in an interview situation, the `new' circumstance would have
11600	to involve some new type of linguistic interaction which the model is
11700	capable  of  responding  to. From its behavior one might then predict
11800	how paranoid patients would behave under similar circumstances.   The
11900	requisite  empirical  tests  and  measures  would  show the degree of
12000	correspondence between patient and model behaviors.
12100		This  possibility  is  of importance in considering therapies
12200	for patients tangled in the quandaries of the paranoid mode.    Since
12300	the  model  operates  at  a  symbol  processing  level  using natural
12400	language, it is at this level at which linguistic and semantic skills
12500	of  clinicians  can be applied. Language-based or semantic techniques
12600	do not seem very effective in the psychoses but they  are  useful  in
12700	states  of  lesser severity. A wide range of new semantic techniques,
12800	including extremes,  could  be  tried  first  on  the  model  without
12900	subjecting patients to blind experimentation.
13000		While our group has used the model to explore a theory and to
13100	study  psychiatric judgements, its potential use as a training device
13200	has not escaped our  attention.   Medical  students  and  psychiatric
13300	residents need `disposable' patients to practice on without jeopardy.
13400	The paranoid model can print out a trace of its inner  states  during
13500	and after an interview.   Whether the optimal  goal  of  interviewing
13600	(gathering  relevant  information without upsetting the patient), has
13700	been achieved, thus can be estimated.  A  beginning  interviewer  can
13800	practice  in private or with a supervisor present.  Many interviewers
13900	have reported that the model has  a  definite  effect  on  them.  The
14000	student  can  get  the  feel  of  the  paranoid  mode  long before he
14100	interviews an actual patient.   The effect of various interviewing
14200	styles can be studied and compared. 
14300	.SS Extra-evidential Support
14400		Besides  the function of evidence in evaluating models, there
14500	is is the role of extra-evidential support. This support derives from
14600	plausibility   arguments,   comparisons   with  competing  models  or
14700	theories, and coherence with other domain theories.
14800		A theoretical model is evaluated relative to viable  rival
14900	explanations. The model I have described stands as a candidate for  a
15000	consensus   explanation   of  paranoid  processes.   It  has  greater
15100	explanatory power than ,say, the "homosexual-conflict" theory because
15200	it covers a more comphrehensive range of facts. The expert forum will
15300	decide its ultimate status. 
15400	
15500	.SS Improvements
15600	
15700		As  the  validation  studies indicated , further work must be
15800	carried out to improve the model's linguistic  limitations.   A  more
15900	sophisticated   language   analyzer  is  being  developed  by  Horace
16000	Enea.(See Enea and Colby  [  ]).  So  many  special  cases  arise  in
16100	unrestricted  natural-language dialogues that a large number of rules
16200	must be written to cover them.
16300		Future versions of the model will  have  a  greatly  expanded
16400	data-base  or  "memory"  which  will  allow  it to answer a many more
16500	questions than the version  described  herein.  The  memory  must  be
16600	interfaced  with  the  language-analyzer in such a way that when some
16700	input expression is not well understood, the memory can still attempt
16800	an  appropriate answer. This capability is especially needed when the
16900	interviewer  refers  back  to  something  already  mentioned  in  the
17000	interview.
17100		Another  improvement  involves  the  ability  of the model to
17200	build up a model  of  the  interviewer  as  the  interview  proceeds.
17300	Information  from  this  submodel can then be used in contending with
17400	the interviewer who is believed to be malevolent. Instead  of  making
17500	rather primitive general remarks about the interviewer the model will
17600	be able to be much more specific in its comments.
17700		Finally the theory might be extended in two ways.  First  the
17800	model  could  be  made  more dynamic over time. The version described
17900	here changes only over the course of a single interview.  To  explore
18000	how  changes  can  be brought about through external input, the model
18100	must have capabilities for self-modification over longer  periods  of
18200	time  in  which it interacts with a number of interviewers. A further
18300	use of more dynamic models will be  to  explore  the  ontogenesis  of
18400	paranoidness, that is, how a system grows to be the way it is through
18500	socialization.
18600		A second extension of the theory would involve  the  addition
18700	of postulates to account for additional properties such as arrogance,
18800	contemptuousness, and grandeur which are often found associated  with
18900	malevolence  convictions.  Implementation  and  integration  of these
19000	postulates  in  the  model  would  complexify  it  to  increase   its
19100	comphrehensiveness and scope by extending its repertoire of ethogenic
19200	powers. In widening the scope of a simulation  one  thus  increases
19300	its  explanatory  power  in  covering  a  greater  range of facts but
19400	accuracy should remain a more fundamental desideratum than range.
19500		It should be clear by now that this simulation  of  paranoia,
19600	while  circumscribed in what it attempts to explain, covers a variety
19700	of facts. The explanation is local and restricted in that it accounts
19800	for  only  one type of symbol-processing resource.   Past attempts at
19900	grand scale explanations of all mental processes in all contexts have
20000	failed.   We  need to build up gradually one circumscribed and tested
20100	theory  or  model  at  a  time  so  that  the  field   can   evolve
20200	gradually and soundly and grounded knowledge can accumulate.