COMPUTER SIMULATION OF HUMAN BEHAVIOR, Slides of Decision Making

Download COMPUTER SIMULATION OF HUMAN BEHAVIOR and more Slides Decision Making in PDF only on Docsity!

'*

COMPUTER SIMULATION OF HUMAN BEHAVIOR

E. A. Feigenbaum

May (^1964) '

\

«

P-

\

»

COMPUTER SIMULATION OF^ HUMAN BEHAVIOR

E. A. Feigenbaum

Consultant to The RAND Corporation, Santa Monica, California

Although the emphasis of this paper is human cognitive

behavior (learning and problem- solving processes), a few

remarks about the "simulator" of (^) this behavior, the digital

computer, are^ in order.

Computers (^) as (^) Information Processors

Ordinarily, we think of the digital computer as a very

high-speed number manipulator- -a calculator. Certainly,

the very (^) term "computer" (^) implies numerical calculation;

and indeed, a computer is mainly used for just this purpose

large-scale numerical computation.^ But the symbols of the

computer, the strings of bits it manipulates,^ do not neces-

sarily have to be assigned a numerical interpretation. We

can consider a particular string of bits as a binary number

and manipulate it by means of hardware that performs addi-

tion, subtraction, etc. But^ we^ can^ also give other-than-

numerical interpretations^ to these symbols.^ A^ digital com-

puter is^ a^ general^ information^ processing system, a^ general

• (^) Any (^) views expressed (^) in this paper are (^) those of (^) the

author. They should not be interpreted as reflecting the

views or^ opinions^ of The RAND^ Corporation or^ the official

opinion or policy of any of its governmental or private

research sponsors.^ Papers^ are^ reproduced^ by The RAND

Corporation as^ a^ courtesy to members of its staff.

This paper is based on a survey lecture given periodically by the author to interested audiences.

there is much work being done on modeling of physiological

processes (^) of the (^) brain. However, the focus in this dis- cussion is on (^) the attempt to (^) model processes (^) of human information (^) processing at the psychological (^) level. The

computer is used as an information processing tool for

working out the remote consequences (or implications) of

a complex set of information processing hypotheses. In

the particular^ research (^) to be (^) described, no comparisons

are made between computer organization and neural organi-

zation (^) at the (^) level of (^) hardware (or (^) "wetware," if you will)

although in the popular press one often finds comparisons

of this kind, the comparisons are usually completely wrong

and misleading, and are rarely made by competent research

people in the field.

On a level of integration above the level of either

computer organization or (^) neural organization, we postulate

a level of elementary information processing, or elementary

symbol manipulation. (^) This is the so-called information

processing level of constructing theories of cognitive

processes. (^) This is a (^) very gross and macroscopic (^) level

of processing compared with processing at the level of

neural firings. Elementary symbol manipulations deal,

for example, with storage of symbols, comparisons of two

symbols, decision-making action based on the presence or

absence of some^ symbol, etc. (^) These elementary information

processes are organized as programs and information struc-

tures which, it is hypothesized, are models of processes

and (^) structures used by the human mechanism. The processes and (^) structures form a (^) closed model, capable (^) of being (^) "run"

%

<

»

in a computer. We^ use^ this^ closed^ model^ as^ a^ subject^ in simulated psychological^ experiments.

A stream of^ behavior^ emerges^ from^ the^ computer^ simu-

lation. This^ is^ the^ model's^ prediction^ for^ the^ experiment simulated. From^ experiments^ with^ humans^ we^ have^ a^ stream of observed^ behavior;^ i.e.,^ behavior^ that^ has^ been^ evoked

when we^ have^ put^ subjects^ through^ the^ same^ laboratory

experiments as those^ simulated.^ The^ human^ behavior^ and the model's^ behavior^ are^ compared^ for^ purposes^ of^ valida-

tion of^ the^ model.^ If^ the^ model^ is^ adequate^ there^ will

be a^ great deal^ of^ similarity^ between^ the^ two^ streams^ of behavior. If there^ are^ important^ differences,^ an^ attempt

is made to^ discover^ what^ is^ wrong^ or^ missing^ in^ the^ model.

Often this loop^ must^ be^ traversed^ many^ times^ before^ a

model is^ developed^ in^ which^ confidence^ can^ be^ placed.

To accomplish these^ modeling^ tasks,^ powerful^ informa-

tion processing^ computer^ languages^ are^ needed,^ and^ some

have been developed.^ Perhaps^ the^ best^ known^ of^ these^ is

IPL-V, the language in^ which^ most^ of^ the^ work^ on^ the

simulation of^ cognitive^ processes^ has^ been^ done.^ Two different types of^ simulation models^ have^ been,^ and^ are being, constructed.

Computer Simulation Within^ Existing^ Frameworks

On the^ one^ hand,^ some^ investigators^ have^ been^ inter-

ested in casting already-^ existing^ theories^ into^ an^ informa

tion processing^ framework^ to^ test^ the^ sufficiency^ and^ the

implications of these theories.^ For^ example,^ a^ theory^ of neural cell assemblies was^ tested^ at^ the^ IBM^ Research

Center, by Rochester, Holland,^ Haibt,^ and^ Duda,^ around

«

New Information Processing^ Models

The (^) second class of^ models^ that^ have^ been^ constructed comprises new and^ different^ models^ of^ cognitive^ processes-

models not previously^ proposed^ in^ some^ other^ form.^ For^ ex-

ample, (^) one such model is^ of human verbal^ learning^ processes,

worked out by the^ author^ and^ Professor^ H.^ A.^ Simon.

It is a^ model^ of^ the^ acquisition^ of^ verbal^ material^ by

subjects in verbal^ learning^ experiments;^ e.g.,^ acquisition

of nonsense^ syllables^ in^ serial^ list^ and^ paired-associate

presentations. This model^ can^ be^ exercised^ in^ rote^ learn-

ing experiments^ of^ many^ different^ kinds,^ some^ standard and some not so standard.

The hypotheses^ which^ comprise^ the^ model^ are^ realized

as programs^ for^ performance^ in^ a^ verbal^ learning^ situation

(e.g., responding to a^ particular^ stimulus,^ recognizing

particular stimuli),^ and^ for

the discrimination among^ the

learning task;^ familiarizing

matching memorized images^ to

learning (e.g.,^ acquiring

different items ia^ a^ verbal

and integrating the^ items;

detect differences^ between them; associating^ memorized^ images^ of^ the^ external^ stimuli

with each other^ so^ that^ the^ presentation^ of^ one^ will^ cause

the (^) evocation of the other associatively)^. This model,^ called^ EPAM^ for^ Elementary^ Perceiver^ and

Memorizer, has been programmed^ in^ IPL-V.^ It^ consists

of some^2000 instructions^ in^ this^ computer^ language.^ EPAM

has been^ run^ in^ about^200 simulations^ of^ various^ kinds^ of

verbal learning^ experiments.^ Many^ of^ the^ well-known^ and

stable phenomena^ of^ rote^ learning^ in^ serial^ and^ paired-

associate learning^ experiments^ are^ predicted^ by^ this^ model.

j

Some of^ these^ are^ quantitative^ predictions^ as,^ for^ ex- ample, (^) the shape (^) of the serial position curve in

learning (the^ fact that subjects learn end^ items^ before

middle items in a serial list leads to a characteristic

bowed error curve which is very stable)^. The model also

gives some qualitative^ predictions for^ certain phenomena

of verbal learning which have never^ been fully^ quantified

in psychological^ literature^ but^ have^ been^ noted^ as^ im-

portant in verbal^ learning.^ For^ example,^ EPAM^ makes

qualitative predictions of the effects that different

kinds of stimuli have on^ the production of^ various types

of errors^ in^ the subject's^ behavior;^ the^ kinds^ of^ intrusion

errors the subject will make; "oscillation"^ of^ correct

responses once learned, then unlearned, then learned again

in the course^ of the learning of^ a^ single^ list.^ In total,

the EPAM model^ predicts,^ either^ quantitatively^ or^ quali-

tatively, about a^ dozen^ of^ the^ most^ important^ stable phenomena (^) of rote serial learning. Many other individuals are^ actively working in^ computer simulation of^ cognitive processes.^ Newell^ and^ Simon^ at

Carnegie Institute of^ Technology,^ for^ instance,^ are^ at-

tempting to construct a^ theory of human problem^ solving.

Their theory receives its precise^ statement in^ a^ program

called the General Problem Solver,^ or^ GPS.^ In^ very general terms, Newell and Simon^ have been concerned^ with how a problem (^) solver builds and searches a^ hierarchy of

problem- solving goals and subgoals.^ They^ are^ concerned

with processes by means of which the problem solver divides

a complex task into simpler tasks,^ the^ answers^ to which

will take him toward the total solution.

.*

I

\

is going to be, Feldman asks him the^ reason^ for^ his

choice. Here^ are^ examples^ of^ reasons given^ by^ subjects:

"You're going to give me^ u£ next time^ because^ you've been (^) giving me a run of ups"^ ; or, "The next one^ is going

to be^ down^ because^ the^ last^ one^ was^ u£ and^ you^ had^ been

giving me^ a^ run^ of downs and the last one^ was^ just^ to throw me^ off." Feldman' s^ model^ is a^ model^ of^ how^ subjects

construct hypotheses^ about^ the^ sequence^ of^ events^ in^ this

binary series.^ The^ hypotheses^ constitute^ the^ reasons

given by^ the^ subject^ for^ his^ choices^ in^ this^ experiment.

Feldman fits the model to particular^ subjects'^ behavior,

using one^ sequence^ to^ set^ the^ parameters^ for^ the^ particu-

lar subject and using a^ second^ sequence^ to^ validate^ the model of^ the^ subject. Another interesting computer simulation^ is^ that done by^ E.^ B.^ Hunt.^ Hunt's^ model^ deals^ with^ behavior

in a^ standard concept- formation task.^ Subjects^ are^ pre-

sented exemplars^ and^ nonexemplars^ of^ a^ concept^ the^ ex-

perimenter has in mind. After the learning^ phase^ is

over, subjects^ are^ asked^ to^ identify^ future^ presentations

as exemplars or^ nonexemplars^ of^ the^ concept.^ Hunt^ pro-

poses a (^) mechanism that grows classification trees^ for sorting new^ items^ to decide^ whether^ or^ not^ these^ items

fall into one category or^ another^ (i.e.,^ whether^ they

are examples of a certain concept, or^ not).

R. Abelson^ of^ Yale^ University^ has^ developed^ a^ model

of the processes^ used^ by^ humans^ to^ assess^ the^ credibility

of statements prior to^ making the decisions^ to^ believe

these statements.^ (9) ' His^ program^ uses^ a^ large^ memory

k

k.

1

structure in which "things^ of the world" are^ categorized.

It consults^ this^ memory^ structure in^ a^ particular^ way^ to

decide whether a particular^ statement is explicable^ in

terms (^) of what has been (^) previously believed— i. e. ,^ in (^) terms of what has been added to^ its memory^ structure. If^ it can (^) find a (^) path of explication--!. e. ,^ if (^) it can find that

this new statement is explicable in terms of^ statements

already believed— it will assess this belief as credible;

if not, it will reject the statement. Sometimes,^ when

it has strong grounds for^ both^ believing^ and^ not^ believing

this statement, it will^ question^ whether^ or^ not^ the^ state-

ment is credible. G.P.E. Clarkson of the Massachusetts Institute of Technology has developed^ a^ very^ careful^ model^ of^ human

decision-making in trust investment decision-making

situations. Clarkson investigated^ the^ decision-making

behavior of^ a^ particular^ trust investment^ officer^ in^ a

bank in Pittsburgh,^ detailing^ how^ the^ man^ thought^ about

his investment problems. He^ asked him questions,^ observed

what the^ man^ was^ reading,^ asked^ for^ interpretations^ of

the reading,^ gave^ artificial^ problems^ to^ try^ to^ elicit

the man's problem- solving techniques and^ decision-making

heuristics, and so^ on. Clarkson then wrote a^ computer

simulation of the man's decision-making^ processes.^ In^ a test of this model, done^ after the first^ quarter^ of^ the following year, when the bank officer had four^ different

portfolios to select, Clarkson used his program^ to^ simulate

the selections, using up-to-date information^ on^ client,

amount, stock market, and^ so^ on.^ In^ all^ four^ cases,^ the

(

V

It

have described is^ the information^ processing^ model,^ not the computing machine. The^ computer^ simulation^ is^ used

as a^ means of^ realizing^ the^ information^ processing^ model

in a precise form, exactly^ as^ a^ physicist^ would^ realize his theory formally^ in^ mathematical^ notation.^ Mathematics

is not a very^ convenient^ language^ in^ which^ to^ state^ these

information processing^ theories.^ Computer^ language^ is convenient, not only^ because^ it is^ designed^ to^ handle information processing,^ but^ also^ because^ once^ a^ model^ is

expressed in this language, a^ computer^ is^ available^ as^ a

tireless tool for generating^ remote^ consequences^ of^ com-

plex sets of information processing^ hypotheses.

r-

I

v

♦

»"

REFERENCES

1. Newell, A., F. M. Tonge, E. A. Feigenbaum, B. F. Green,

Jr., and^ G. H. Mealy, Information Processing Language-

V (^) Manual , (^) 2nd cd. , (^) Prentice-Hall, Inc., Englewood Cliffs, New^ Jersey,^ 1964.

2. (^) Rochester, (^) N. , (^) J. (^) H. (^) Holland, L. H. Haibt, (^) and W. L. Duda, "Test on a (^) Cell Assembly Theory of the (^) Action

of the Brain, Using a Large Digital Computer," IRE

Trans. (^) Info. Theory,^ IT- 2 (1956) (^) 80-93.

3. Colby,^ X. ,^ "Computer (^) Simulation of a Neurotic (^) Patient,'

in Computer Simulation of Personality, Tompkins and

Messick (eds.), (^) Wiley, New (^) York, (^) 1963.

4. (^) Gullahorn, John, and Jeanne (^) Gullahorn, "A (^) Model of Elementary (^) Social (^) Behavior," in Computers (^) and Thought,

E. A. Feigenbaum and J. Feldman (eds.), McGraw-Hill,

New (^) York, (^) 1963.

5. Feigenbaum, E. (^) A., "The Simulation of Verbal (^) Learning Behavior," in Computers (^) and Thought, (^) E. A. (^) Feigenbaum and (^) J. Feldman (eds.), (^) McGraw-Hill, New (^) York, (^) 1963.

Also published in Proceedings of the Western Joint

Computer Conference (1961) , Institute of Radio

Engineers, New York, 1961, (^) pp. (^) 121-132.

6. Newell, (^) A., and (^) H. A. Simon, (^) "GPS, A Program that

Simulates Human Thought,"^ in Computers and Thought,

E. A. Feigenbaum (^) and J. Feldman (eds.), (^) McGraw-Hill, New York, (^) 1963. Also (^) published in Lernende Automaten, H. Billings (cd.),^ Oldenbourg, Munich, (^) 1961.

7. Feldman, J., "Simulation of Behavior in the Binary

Choice Experiment," in Computers and Thought, E. A.

Feigenbaum and J. Feldman (eds.), McGraw-Hill,

New York, 1963. Also published in Proceedings of

the (^) Western Joint Computer (^) Conference (1961) , Institute of Radio (^) Engineers, New (^) York, (^) 1961, pp. 133-144.

8. Hunt, E. 8. ,^ and (^) C. I. Hovland, "Programming (^) a Model of (^) Concept (^) Formulation," in Computers (^) and Thought,

E. A. Feigenbaum and J. Feldman (eds.), McGraw-Hill,

New (^) York, (^) 1963. Also published in Proceedings of the (^) Western Joint Computer (^) Conference (1961) ,

Institute of Radio Engineers, New York, 1961,

pp. 145-155.