draft-unified theories of cognition-ch3-human cognitive architecture, 1987-allen newell

10
8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 1/10 DRAFT  MAT ER IA L : LIMITED DIS TR IB U TIO N F OR COM MENT NOT F OR QU OT ATI ON File: C3H CA. MSS Current version: 4 Aug ust 87  17:44 Started: 25 Jul 87 The  1 987  Willi am James Lectures u UN IFIED TH EO R IE S  OF CO G NI TI ON CH A PTE R 3 HUM AN CO GN I TIV A RCH I TEC TUR E D RA FT  Alien Newell 4 August 1 98 7 Departments of Computer Sc ien ce and Ps ycholo gy Carnegie Mellon University Pittsburgh Pennsylvania 15213 W ° £  / ' b^ 'fi 3

Upload: contisoldat89

Post on 07-Aug-2018

217 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 1/10

DRAFT

 M ATER

IA L:

LIMITE

D DISTR

IB UTIO

N

F

OR

COM

MENT 

NOT FOR

QUOT

ATION

File:

C3H

CA.M SS

C

urrent

version

:  4 A ug

ust 87

 

1

7:44

Started

:

25 Jul

87

The 1

987 

Will

iam Ja

mes L

ectu

res

u

UN

IFIED

 T

H EO

RIE S

 OF

 C OG

NITI

ON

C H A

PTE

R

3

HUM

AN C O

GNI

TIV E

 A

RC H

ITEC

TUR

E

D

RA FT

 

Alien

 Newe

ll

4 August 1

98 7

D

epartm

ents

ofCom

puter

 Scien

ce and Ps

ychol

ogy

C

arneg

ie Me

llon U

niver

sity

Pitts

burgh

  Pe

nnsylv

ania 

15213

W° £ 

A

 

/

'

b^

'f

i

3

Page 2: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 2/10

Tabl

e of

on

t nts

1

. The hum

an is a

 

sy

mbol

 sy

stem  

2

2. Sys

tem Lev

els

4

3. T

he Tim

escale

 o f

 

Hum

an Act

ion

6

4 .

 T

he

N

eural

 B and

 

7

5.

Neura

l Cir

cuit Level

8

6. 

The R

eal-Ti

m e

ons

traint

on  

C

og nitio

n

10

7 .

The Cognitive

Band

12

8.

Level

 

o

f Simpl

e Op

eratio

ns 

16

9

. Lev

el of G

en era

l Op

eratio

ns

1

0 . T

he

I

ntend e

dly Ra

tional

Band

 

2

1

11. High

er  Ba

nds  M So

cial

H

istoric

al and E

voluti

onary

 

2

2

12

. Sum

mary

 

24

W

J Ch

. 3. H

CA : Prelim

inary dr

aft

of

0

4

A ug

ust 87 17 :

44. Lim

ited distri

bution

  D

o n

ot quote .

Page 3: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 3/10

 

List

 of Fi

gures

Figure

 

2-1:

Exp

ansion of

sp a

ce 

w

ith

le

vels.

5

Fig

ure 2-2:

Ex

pansion of

space w

ith le

vels.

6

Figure

3-1:

Ti

m escale

o

f human

 ac

tion.

7

Figu

re 4-1:

The

neu

ral

level

.

8

Figure

5 -1 :

The neura

l c

ircuit

leve

l.

9

Figure 7 -1:

The 

necessar

y phases

 

o

f delib

eration.

1

3

Figure

7 -2 :

Properties

for

au tomatic

and

contro lled behavior.

14

Figu

re 7-3:

  Ex

ample of aut

om atic

an d c

ontrolle

d behav

ior (Shiffri

n & Sch

neider,

1977) .

15

Figur

e 8-1

: The four

 levels

of the

cogniti

ve

ban

d.

16

F

igure 9 -

1:

Compre

ssion of 

the level sc

ale factor

to s

queeze N+l le v

els 

into

N

.

1

7

Figu

re

9 -2

:

Resid

enc e times for

variou

s tasks

 

(Sim

oN 7 2).

18

Fig

ure

9

-3 : 

19

Figu

re 9 -4

:

20

Fig

ure 10-1:

The inte

ndedly 

ratio

nal

ban

d M  

knowled

ge-level

  systems

.

21

F

igure 11 -1

: Hig

her ba n

ds.

22

W

J

Ch

. 3. H CA

: Prelim

inary dr f

t

of

0

4

A ugust

87  

17:44

. Limited

 d

is tributio

n. D o not quote

.

Page 4: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 4/10

Un

ified Theorie

s

of Cognition

Chapte

r 3. Human

Co

gnitive

A

rchitecture

 

/

 

In this

lecture w e turn

to the hum an cogn

itive architect.ure.

iS j~lhe  last

lecture >

f c provided o u f 3 e , K F S   T m J .L h u the

basic concepts necessary

to

understanding

v

>

intelligent

systeiy^

Representation, know ledge,

symbols .and sear

ch

pply

as m uch

to machines  and to

 

hum

ans.  H u m a n

s,

of

  co urse,

are

sp

ecific  in n^ny

ways.

Back

in

  Figure 

U

TC- CO N STR A T

N TS

we laid ou t ma ny

of the co nstraints

that shape the

human

mi

nd.-k

For example,

d t

they are

co

nstructed of neural te

chnology, tiwithey

arose through ev

olution, and tip*

 

th

ey m ust

be

high

ly autonomous

.

 

a

theo

ry

b

rms

t

he

Our ult imate goal is

a unified th eory

of

hu

ma n cognition.

That w ill

be

ex

pressed, we h

ave maintai

n^  

of

  the architectu

re

of

human cog n

ition hat is, of

the fixed  (or slowly

varying)

structure

framework for the

immediate  pr o

cesses

of.-c

ognitive

perform an

ce and learning .

  Tjbus, w e ne

ed

to set

out th at

 Q

\jA<Qi.c

architectur

e. We wil l do so in

phases.  In

th^

lec

ture w e will

attempt to derive

some.aspects

 

o

f

the

human cognit

ive

architectu r

e, at tending only to

the way  the human is

situated

  in

the

world. This

will lay

the groundw or k

  for

proposing in the

fourth lecture

a

sp

ecific architec

ture in

d

etail. (T

hough, ev en there, vty

io uc aspe

cts will re ma in

e

.Alternative^ , the

architecture   c

ould

be

laid as  

a

total

system. ifclose/£hpp

roximati*l4e

the axi o mat

ic

ideal.

A n 

a

dvantage of   the

phased description

s, in addition to

whatever didactic vi

rtues it m

ight

ha

ve, is ^ trf separa

tu» w hat

can

be claimed ab o

ut

t

he architecture on t iiKg

y gene

ral grounds, f

rom what

must be justif

ied

by de t

ailed

r i v

f tr

ex pe

rimental da t

a}  

/Thu

s,

this

  lecture will

be devoted to

a quite gen eral

argument ^Hw^tf

ill start fr om the neural

 

technolog

y, which is   quit

e clearly

th

e techn

ology of   the human

  cogn itiv

e

ar c

hitecture.

This architectu r

e

m u

st

su pport

mind-like behavi

or. From th e

prior lecture w e alread

y

have a

 ch

aracterization of

what that

m

ea ns. W

e

will

add

to thes

e w

hat we will call

the

real

-time constrain

t on human

cognit ion. From   this th r

ee constraints,

w e

w

ill

deriv

e

a

number of ke

y

asp

ects

of

 

th e

 cog

nitive architecture.

H

owever, w hatev

er we are able to

derive from   them

^te

no

lo nger

 

o

he ila a

re indeed

rather

generalt

 

options in

specifying  the

details

o

f the h u m a

n architecture.

The wisdom

of

such

  a strategy

f a ttempting

  to

divide and con quer hould be evident f rom the discussion

in  

the last lecture. T here (F igure FC S -A R C H V A R ) w e

saw, thai

large degre

egofj^egdgrn \v£fe

available to

 construct arch itectu

res

of

symbolic system

s. W e also

noted th e

the archit

ecture

may

be t t e essential

ly^hidden va

riable

hat there is

n

o  way to

determine

wha t'represen

ta t ions^?  : whjg gg m

trol structure

s are

u

se d. Since

any general (i.e.,

universal)

architecture can mimic

any other,

the si tuation is

hopeless.^lear

ly

then,

w ha

tever

a

spects of   th e

 architecture can be

pinned down from the

general situatio

n within whic

h

t

he huW in ar

chitecture is  co n

structed and operat

es

uc

h  can only

be of g

reat

help

in

m aking

 the internal de

tails identifiablksspn

e might have some

doubts

about whether an

ything

c

an be 

s

aid, but

as

^ ~

 

w e

will

seem

 im

mediately

that

is

far from

so 7

~-^4 Q

U J »

H er

e is anoth

er

way

to think a b

out the enterprise of this le

cture .

Evolution is th e

designer of th e h um

an cognitive 

ar

chitecture.  It wil

l pick and choos

e

s

ystems that wi

ll aid

in

th e survival

 of the species.

Our pr oblem as sc

ientists is

to guess what des ign evolution has settled for to date. W e know a little about evolution

as

a designer. It never

starts

over, a

lways working w it

h what

i

s  available.

  In Ja

cob s (1914)

 

now famous

phrase,

evolution

is

a

 

tinkerer.

For

inst

ance, once

species is

co mm it

ted

to

a k-strategy (heavy

  investmen

t

in few

pr oge n

y) evolution will not

shift to

the opp o

site R-strategy (l i

ght investmen t in many

 progeny)

t

is

 

too

hard to

get

fr

om here to

 

there.

Evolutio

n

picks its

^s fotonJVithin th e d

esign constraints

posed

b

y the situation

  in which

the organism finds itself.

If w e ca n

dCL

a

understand th ese

constraints, then we can

 so»m

oro

cloatly

tho fimitpjl fi e

ld within which

evolution  

must

operate.

T his

lecture

will

 b

e

fa i

rly speculative. Any

attempt

 

to

get at genera l ch

aracteristicsA must

run

t

his

ri sk

. W

e

will

 

WJ

Ch. 3. HCA:

 

Prelim

inary draft

of 04 Aug

ust

87

 

17:

44. L

imited distributio n

.

D o

not

q

uote .

Page 5: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 5/10

The

more s

ubstantial argume

nt is

to

ref

lect

on the varie

ty of re

sponse fu

nctions th

at humans

 

  ot

can do,

but

do

. The

fa

ct

is

that

the y

seem

  to create new respo

nse fu

nctions

all

th

e,

time.

To

 adopt

the w

ell-worn 

device of

the

 .lio*-. £

Martian biologist, 

wh at

would impress

him

most

if

he   looked

at

the human species  biolog ically

is

'attem

pt to

 asses

s

th

at risk

 

at

the end

of th

le

cture, after

the results

are  befo

re

us .

Even

 if

the spe

culations are

wro

ng,

I

th

ink/they ¥*H be

 wro

ng inkitoro o

ting^ays

.

i

c *-

Aju

^-aJUx^S

Tx

H

ere is a preview

 

of tne resu

lts. Dif

ferent cogn

itive world

s have different

-time

sca

les.

The

different

 kin

ds

o

f

cognit

ive worlds

 

that

we

 see

a

re governed

by 

the time  s

cales on 

which

thv occur

he ne

urologica

l,   cognitiv

e,

ratio

nal, and social.

Th

ey are wh

atthey are t

o enable

  mind-lik

e

beh

av ior

to em

erge o get

to co

mputation

al ,

—————— —

—————— —

— »^

j <

tr»/«*4t

^

symbolic

sy st ems/^

W ithin

the cognitive

werid

there 

must be a

distinction

between

au tomatic and controlled

process

es ,

which aga

in

is

a d

istinction 

in

le

vel.

In additio

n, the arc

hitecture

has to

 

be recogn

ition-base

d and

 the

re

ha

s-to be a cont

inual

shift to

wards inc

reased re

cognition.

  This cor

responds to

continu

al

mov

em ent alon

g

the isoba

in Figure FCS-PRE

PD ELffiT

RAD EO

FF, t

owards in

creased p

reparation

.

1 T

he hu m a

n is a

  s y m b o

l syste

m

The treatment

in

t f r e last le

cture

wa s abstract

 

in a

very particu

la r 

way. I t

 

discusse

symbol-le

vel system

s and

kn owledg

e- level

system

s as the

 general str

ucture that

wa s

 

nece

ssary to ob

tain gen

eral inte

lligent beh

avior. I d

id not

spec

ifically make

the case tha

t h

umans Jw

erefoiow4

63ge systems a

nd sym

bol sys

tems

hough

it

w

as clear l

y

understo

od that wa s

what we

were after. *3it

we

settled

for underst

anding

the nature

of these

various

s

ys temf

and

what gay^

rise to them,

namely

 the need

 

to

 

d

eal with

large am

ounts o

f variabilit

i

n respo

nse functions

,

i

A

t this

 

p

oint 

we wish t

o

be

ex

plicit that hum

ans

are symbol sy

stems th

at are at least m

odest

appro

ximations

 

of

knowle

dg e sy

stems. 

They might b

e o

ther kinds

 

of

systems

 as w

el l, but at least

they ar

e symb

ol

s

ystems.

The

ground

s for thi

sjargumen

t, as made

 

clear

by

the entire

previous lec

ture,

is the variety

 of resp

onse functi

ons

tha

t the

human use

s. Ifjy|ri£

ty of

response

 

f

unctions

is

i

mmense

enough, a sy

stem wil

l

be

 dri

ven t

o

compose

 its 

respon

se 

functions

 

b

y

m

eans of

a

ccono

utational

system that 

construc

ts

repr

esentation

s b

y me

ans of

co mposed

 

trans

formation

s and^use

^ symbols

 to obtain

distal acces

s. That  w

hole a

pparatus ex

ists

be

cause o

f the  d

emands of

varie

ty. The groundw

ork

is laid. W e j

ust wi

sh to

 

be

explicit ab

out its ap pl

ication to

humans.

O

ne weak

argument

is that

human

s can un

doubtedly

 

emulat

e a

universa

l

madiin

e.

They

 

migh

t do  

it

rat

her

slowly, becaus

e th

ey may ha

ve to sp

end

a

 large

amou

nt

of

time m

emorizin

g tflt

new 

st

ates.

B ut, if we  

wait long 

enough, they can

do perf

orm

t

he op

erations of a

univer

sal machin

e. They are

 of

course lim

ited in

 

their

lifetime

s

(measu

red in terms

of total numb

er

o

operations

 they

 

ca

n perform) 

and ulti

mately

in the

reliab

il ity o

f their me

mory. 

B

ut

ne i

ther

of

th

ese

is of the

essenc

e,  

j

ust as thejta

re

n

ot

for Q O

j f l r j u t e r s w

hich also 

have li

mited lifeti

mes and

reliab

ilities. T

hus, techn

ically they

 

a

re the

 kind of 4beast

vv f

cSncan

be

universal

 

m

achine. 

AfclW

th

that

comes

,

^

v

- ag ain

  technic

ally, a

ll ther>ther pro

perties. Ho

wever,

the argum

ent is we ak, bec

au se t

he ima

gined type 

of verifica

tion

'

^ sp>

oifying

to a hu maJra

 spec

ific univ

ersal machine

f C s a y a

Turing

 

M

achine) a

nd then

observ

ing the pe rson's

executi

on of

it

jn terpretiv

ely

ao to op

cak s ar

tificial, n

 is

n

ot

an observ

ation on th

e

ki

nd o

f life

that

 hum

ans lead.

TT c

ould b

e the e

xploitatio

of a

capab

ility tha

t is actual

ly irrele

vant to

 

th

regul

ar style

  with 

which th

e

hu

man

interacts

 

w

ith ^

Environm

ent.

eff l

orescence

 

of

adaptat

ion

Humans

 

appear

 

to  go aroun

d simply 

creating 

opportu

nities

of

all

kinds

  to

  build

diff

erent resp

onse

function

s.

L

ook

at the varie

ty

o

f jobs in

the

 

wor

ld. Eac

h one h

as humans do

ing

dif

fer ent

kinds of

4 JkA W n

O

response fun

ctions. Hu

mans invent

games.

They ha

ve a

ll different kin

ds

o

f s

ports. They

 

n

o soone

r invent one,

than

the y

invent new

  ones.

They not onl

y inven

t card game

s, t

hey 

collect

 

them in a book

and publish

thejj* /

(15

0 strong)

(H

oylSO).

Th at

implies that p

eople b

uy

the m so

they 

ca

n develo

p

y

et new

respon

se function

s

by^ore

. .

They also

 

dance, w

rite

bo

ok s,

have c

onversatio

ns. A c

onversati

on

is

not

hing but a

n oppor

tunity to in

teract) w

ith t

he

I

W J C

h 3  H

C A : Prelim

inary draft

o  0

4

A u

gu st

87 17:44 Limit

ed distr

ib ution

D o  not q

u ote

Page 6: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 6/10

 

envi

ro nmen t

 in a

way th at

w

as

differe

nt th a

n

p

ri or

interac

tions hat

is, t

o

bu

ild

new

r

esponse

 fu n

ctions. Think

 

o

the L ib

rary

of

C

ongress

 as eviden

ce of th

e varietyof

resp

onse func

tions

t

ha t h

uman h

ave

e

xhibited

 by

 w riting th

em,

will exhibit

 

by

read

in g

them

, and

w

ant

to

e

xh ibit by

  building

 b

uildings

 

t

o make

 them

 

ava

ilable. To

ac ad

emics

the

mentio

n of bo oks

sugg

ests in

tellectua

f

unctions

, as if

all this w a

s perh

aps

a

phenom

en a of

th e 

high en

d

of

th e 

so

cio-eco

nomic scale.

W hat t

hen of

rapping

? A

 

cr ea

tion

o

the

b

la ck

ghetto ,

rapping

 is an in

vention

 to produ

ce an

op

portunit

fo r

 ne w  respon

ses

to

a highly dyn

amic en

vironm e

nt. It ser v

es

o

th er

funct

ions

as

w

ell,

bu

t i t

build

s on

t

he hu

m an

pr

oclivity

to inve

nt ne

w form

s of

 

re s

ponses.  Peop

le  sa^ £

create

th

ese opp

ortunitie

s. Indeed,

our

M

artian biolo

gist  woul

d

not

be

 w ro

ng to

conclud

e  that

th e

bigg

est biolo

gical pu

zz le abou

earthlin

gs is w hy t

hey

hav

e dev

eloped this efflo

re scence

 of

adaptat

ion.  

 

Etholo

gy,

lo

oking

at other

orga

nisms

from

d

ig ger

wasps

 to herring

  gulls, 

has pro

perly

b

ec ome

in

trigued

w ith th

e

a

daptatio

ns they ex

hibit. E ach

ad

aptation

 

is /a

 uniq

ue biolo

g ical phe

nomena,

  eachis

to

be

understo

od  b y

ex p

loring

the

 

be h

avioral

an

d

phys

io logica

l m

echanis

ms that s

upport it.

Th

ey

ar

e to

 

cu

rated on

e

by

o

ne . Not so

 

w

ith huma

ns/

T he

re is n

o

enumer

ating th

eir 

adaptat

ions

hey w

il l 

inven

t n

ew ada

ptations faster

th a

n they can

 

be

 re

cord^ The

act of  recor

di ng

is itself

 one m ore

 ad ap

tation

r rather a

 w ho

le

ge

nerator

of 

them,

as the

 

pr

oblems of

suc

h

sc ie

ntific enterp

ri se

unfold

 and are

 respond

ed

to .

W hat

I am

sa ying

is

not

new,

nor

it

is

supposed

 to be

. There are

  manyw

ays

o

talking abou

t t

he life

o

f  h

omo

sapie

ns, so as

to reve

al

what

 is th

e life of

the mind.

A lw a

ys

eo

oe ntiolly

 t

he

s

ame

gross facts

 

ar

e to be

describ

ed. Th

e

iss

ue

is

w hat

 

it

t

akes to

convinc

e ourselve

s

that

 hum

ans

deal w

ith sufficie

nt variet

y , so

t

ha t 

they m

ust

be

 sy m

bol

sys tems

hat no

system  

of less powe

r a

nd unive

rsality c

ould suffice.

 

I

a

m

atte

m pting

a d

escripti

on

th a

t take

s as

th e given

s obser

vations on

  the

b

ehavior

of

  human

s, to w i

t, o

n the variety

 

of

th a

t behav

ior.

I

w

ish

to avoid

t •ysfr '

 

ju

dgm ents

 on  th

e

c

ontent of

  th at

beha

vior, fo

e

xample,

  onits

rationa

li ty

or

 o

eg roe of

its

a

daptatio

nfTesp e

cially

w

ish to

 

avoid

 invol

ve m ent

w i

th  an y

 

in te r

nal or

  structur

al aspect

s of th

hu

man. 

M y o

bjective

 is to g

ro und

th e 

assert

ion that th

e hum

an is a s

ymbolfs

ys tem

on

exte

rn aLasp

ec ts,

so th at

it ca n 

serve a

s

a

des i

gn con

s traint f

o r

con

sidering

 

the

 

nat

ur e of

th

e internal stru

cture. 

~ J

LoJ^oJ

^

A

/

u n J L

To in dicate

th

is

beh

av ioral

ch aracte

r of

humans,

 

I

 use  

the ph

rase

un

limited q

ualitative

 a

daptatio

n. C l

early,

h

umans  a

re not

infinitel

y adaptive

. That i s

easy to

sho

w .

Ju s

t

p

it 

two hu

m ans against

each o t h e

B J i n a

compe

tition;

one

w ill

w in,

th e

other

w ill lose. The human

th at

loses clearly was

not

su ff iciently adaptive.

Th^

it is

the variety

of 

adap

tations th at

is

a

ss erted

heir ra

nge^nial

it atively spe a

king. 

Here th

ejj seem s

to

b

e no

  limit wh

atsoever

.

W

hat

m ight

ha

ve seem

ed

a

l imit

, in term

s

of

sp ec

ific sen

sors an

d time-b

ound con

tact, h

umans m

anage

to transce

nd

by

 

instrum

ents and histor

ie sy

They

 

e

ven bu ry

t ime c

apsules

so that so

me fa

r-future

 

h

uman w ill

hav

one

ad d

itional

opport

unity

to

beh

avjfcf adapti

vely

with respe

ct to a p

ast he

 

or

s

he

might

 o

therwise

 

h

ave misse

d.

T

his ar gum e

nt h

as

a s

of t

spot, w hi

ch shoul

d

b

e noted

. I

t is an

 

as y

mptotic

  argumen

t.

T

ha t is

, as

the variety

 

of

function

that

ca n

 

be ex h

ib ited b

y a

system 

increa

se s wi tho

ut

l

imit,

w e know

 

th

e set of

funct

ions

be

comes

the set of

compu

ta ble func

tions.

Furtherm

ore,

w e know such

 a set

ca

n

be

gener

ated only 

by a u

niversal

  computa

tional

system

s (exac

tly so

s

ince th ese

tw o

  notions 

simply g

o together)

. If   w

e

co

nsider

sy s

tems th at

prod

uce an ever

grea

te r

variet

o

f fu n

ctions, at

some poi n

t

th ey must

have

the structu

re

of

a

universa

s

ystem,  i.e.

, of a

symbol

syste

m . A

  huma

n is

ca

pable

of p

roducin

a

imme

nse variet

y

o

f fu n

ctions and d

oes/so

in

 

its  

everyda

y life. B u

t is

I

his ^enou

gh 

variety s

o

th a

t

the structu

re

m

us t be

that of

 symbol sy

stem ? Co

m putatio

na l

theory  do

es no

t

yet

 pro v

ide

any

useful answers

 to

 this

 

qu

estion n

part

because

 it

hasn

 t  sough

t therr

i,afike«g

h

in general 

such

 

q

uestions are

 

hard

 to

an s

w er in

use f

ul w

ay s. It w

ould

b

e n

ice to

have som e

theory 

like that,

  but

I

 d

on t kno

w of

any. It is

O

f cours

e,

iijf

ould puzzle

 a bo ut it ,

he might a

fortiori

know

th e answe

r- because

th e 

art

ian\

would ex hibit

that

  s ame efflo

rescence. B

ut

n

o

metaph

or ca n

 

be

  perfect.

n w

e

W J

 

C

h. 3

. H C A

: Preli

minary d

r ft

of

0

4

Aug u

st

87 17:44. Lim

ited

di str

ibution .

  D o

 

n

ot quote

.

Page 7: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 7/10

instructive

to

observe that

 computers are

in

th

e sa m e situation,

to

c

ompute any computable

function a computer

has

to have 

the

structure

o f

a

symbol sy s

tem.

We co

nstruct computers

so

that^ia

ve this s

tructure.

But

need

we?

I

s the

va riety o f actual

functions  that w e want

to compute

such

tha t we c

ould get

by with some st ructure, perhap

s far

removed f

rom

a

symbol system

s.

It

seems hardly li

kely, but there is not the mathem

atical theory to provi

de more

definitive answ

ers.

^^ ^^^^^S^^^E^^ ^

p

li

 gsa >

 

w e will now take

it

as

established that the architecture

of

human cognition 

is

a symbol system.

2 System   Levels

Let

 us tu r

n to

underst

anding the technolo

gy out of which th

e human architecture is co

nstructed. The first point

is

that, o f ne cessity, intelligent

systems

are

built

up of m ultiple leve

ls

of

systems.

A system level is

a

collection

of

comp

onents, which ar

e

lin

ked

t

ogether

in

some arrang

em ent and w

hich interact, t

hus

producing behavior

 

a t

that

sy stem

  level.

Mult

iple levels means

 

that

the components at

one level are realized

by system s at

the

next

level

belo

w .

We have o f cour

se bee n here before,

 as witn

ee eed by

Figure FCS

-C O MPSYSHIERARC

HY , which

showed the

computer

sy stems levels. E mpiri

ca lly, everything

w e understand about engineeri

ng  such systems

 

to

get in telligence

is to build up m ultiple levels. This is one

o f

th e great empirical invariances lthough many different w ay s have

been found

to

c

onstruct informati

on processing s

ystems, they sti

ll

all consist

of

a

hierarchy o f l

evels and indeed

essentia lly the sa m

e hierarchy.

I

wish to maintain

th a

t

the

human

 architecture

is

built

up o f

a

hierarchy ha

t

it canno

t be otherwis

e structured

o

that

the

di

scovery of the architectu

re ca n proceed

within this assumption. That eng

in eered computer

systems

see

m

to have a hie

rarchical structure,

as

just

re viewed, c

an

c

ertainly be taken

  as

one piller

of support.

A second

pillar 

comes from

 Herb Simon's analysis fa

r hierarchy (Simo6

2). The

ar

gu m ent there

was

that

stability dictates th

at

systems  have to b

e

hier

archical. To build com pl

icated systems witho

ut first building stable

subassemblies

will

al w ays fail

he entire structure wijl'disintegrate

before it

all gets put to gether. If s

table subassemblies

are

cr e

ated,

layer upon lay

er,

th

en each one A

has

a reaso

nable probability

o f being constructed out o f a

few parts. Thus,

there

e

xists a ge neral argument that

stability dictates

the

ex

istence

levels.

Th

e stability argument,

of aefurse can be

taken

to underlay the en

tire hierarchical s

tructure o f matter, from

 nucleons, to atoms,

to

m ol

ecules, a

n

on up. So the

levels

hypothesis

may have nothing to

 do with intelligent

systems, but simply with

 t h e way all systems are

put together.

Al l we need for

the

argument

 is that intelligen

t system s will be h

ierarchical. i/lu«.

**

fUL^^J j 

^L^-p

^-

 

J C C M I - M , T ^

Levels are

clearly abstractions, bein

g alternative

ways or descrtmng

the same system , each altern

ative, ignoring

some

o

wha t

 is specified at the

level beneath

it. It

is

 all in the head o f

the observer. ^

Ffiere is

more

to it th an th

at.

Levels

ca n be stronger or

weaker depending

on how well the behavior

of the sy stem, as described

at

a

le vel,

canjbe

predicted

or explained by

the structure of

the system described at

the same level. In standard t

reatments^ystems

analysis

(SystemAnalysis),

systems are called state determin

ed  w

hen

their fu t

ure

beh

av ior is determined wnjfl 

th

eir'

2

3

c

urrent s

tate.

That

is what holds for

a strong leve

l.  A

level

is weak if co nsiderations

from

lo

w er levels enter

into

de

termining

the

futur

e course.

In

engineered

systems  (Figur

e

FCS-COM

PSYSHIEARCHY a

gain),  great care

is

taken to make strong  levels o seal

off

ea ch level from the on below. When dealing with logic circuits there

is

no 

need

to

unde

rstand the continuou

s circuitry underling

 them xcept

when things go   wron

g.

W he

n dealing 

with

programs

there is not need to

 

under

stand the register-transfer

 circuits that realize

the ope rations a

nd the

in t

erpreter

gain, except

when things go

wrong. And so

on.

These

are all very str

ong system levels, as evidenced

by how

sm all

the

failure rates

are

in

commercia

lly successful

systems. Many natural

system levels are also

very strong,

such

 

the

atomic and molecular

levels. The stronger a

level

is the

more

it forms

a

dis tinct world,

in which

nothing,

m us t be known about lo

w er  levels

in

order

to live within it.

However,

aH

  natural-sy

stem levels need be stro

ng. In

W J C h  3 H CA:

Preliminary  draft of

04 August 87  17

:44 Limited distribut

ion D o not quote

Page 8: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 8/10

Page 9: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 9/10

 

Figure

2-2: Expansion of space with levels.

with

new

properties. It

can't/be

done with two or three components and it

can

be done with

just

a couple

of

system

times. Novel behavior cantbuild up that

rapidly.

On the other

hand,

there are

lots

of cases where

less than

100

or

1000 components or component-times suffice. Thus, we

can

take a factor of 10 to be the minimum characteristic

factor to get

from

one level to the next. This factor is highly approximate, of course. It might be only 3 in

some

C M i r e m c

cases,

it

might be

as

many as 30.

Let

us

use

0

s a special notation

to

indicate

such

very approximate

numbers. 2

 

The

Timescale of Human Action

Let us now consider Jhe timescale at

which

human action occurs,

as shown in

Figure HCA -TIMESCALE. At

the

o

left

time

is

measure^n seconds, using

a

log

scale.

The next

column

to

the

right

names the

time units, m illiseconds

(ms) to sec/ to mki to hours. The units themselves are, of course, geometrically related to each

other. The

next

n

system with that

characteristic operation

time.

Time is

the useful

measure of system

level,

for us^^ /

space. But of course

the characteristic physical

size

increases correspondingly.

characterized by different theories, and

these

are

shown

in the

right

hand

column.

Different levels are

at

the

bottom, there is

the

neural b a n e } of three

eurons;

and neural circuits,

a factor

of

ten

up .

 

This figure provides an overview of where we

are going. Startin

\ evels eurons: organelles, which are

a

factor of

ten

dpwjvfroL _-__ _,

_._

____ _ __

.

.

•-

Ak/ 4J

jK« «M

<

«•*

TT/t

tiff

A s

we'll*

see, neurons

have

a

characteristic operation

time ofaoout a ms,

and neuraTcircuits

have a

characteristic of

 *+**

a***) ^s^ J>

r\

about

10 ms.

Thpii^here is the

cognitive

band.

Here

the levels are unfamiliar

.r ^ve

called them deliberate acts,

cognitive operajions and unit

tasks.

Each of them

takes

about ten times as long as the

level

beneath. It

will

be the

main

task of

this

lecture to establish these levels

and

something of their criaract^^rx ve^ecpjmitivej?and

lies-the-

the rational band, ^which is of

the

order of

minutes

to

hours.

All of these levels are given

the

same

label,

namely

/as£s.

W e

will see why that is

appropriate. Finally, even

higher

up

thej^ lies

something^alled

the social

band.

which we will

have only a little to say about ts

L

f tm^C?

£*

One striking feature of Figure HCA-TIMESCALE

is

that

each level

is

only the

minimal

factor

of

ten

above its

y v

components.

This

is

evident directly

from

what

we

know empirically about

the

levels

of

the neural band.

Each

new

level

occurs, so

to speak, iarf as

soon

as

it can

in terms

of

components.

3 This justification of the

minimal

property

1 0

rather

than

-10 because we

want to preserve

-10 to

indicate the

usual

degrees

of

approximation. In the

lectures we

used

~

special

character is

not

available

here. 10 is

a place holder for a

special notation.

3 N<ne: This needs work. One

interesting

aspect is

that

what makes for larger

steps

is aridity, as in sys tem plains. On the

contrary,

if one is

trying to get« much complication as possible, then one wants levels as soon as possible.

V

W J C h 3 HC

A:

Preliminary

draft of 04

August

87

17:44 Limited distribution Do not quote

Page 10: Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

8/20/2019 Draft-Unified Theories of Cognition-Ch3-Human Cognitive Architecture, 1987-Allen Newell

http://slidepdf.com/reader/full/draft-unified-theories-of-cognition-ch3-human-cognitive-architecture-1987-allen 10/10

25

Refe

rences

Fodor, J. A .

The

 

M odula

ri ty

o M in

d. C a m b

ridge,

M

A :

Brad

ford Books

, M I

T  Press ,198

3 .

K ol

ers,

P.

A .

M e

m orial c

onseque

nces

of

 

automat

iz ed encod

ing. J

ournal o  Expe

rimenta

l

P sycholo

gy: Hum

an

le a

rning and

m e m or

y,

1975 ,

 1,

689-7 01

.

N ew

ell, A .

Simon, H .

 

A .

H uman

Pro

blem Sol

ving.

E nglew

ood

C l

iffs: 

Prenti

ce-Hall , 197

2.

Posner , M

.

I.

S

nyder, C.

R. R.

Facil i

ta t ion an

d  inhibitio

n

in

 th e

proc

essing

o

f sig n

als. In

R ab

it t,

P.

M . A .

Dornic

,

S. (E

d.), 

Attent

io n

an d

 P e

rforman

ce

V .

N ew

Y o

rk :

A cadem

ic Pre

ss , 1975.

Sc

hneider,

 

W

. Sh

iffrin, R . M .

C ontroll

ed 

a

nd auto

matic hum

an in

formatio

n process

ing: I.

Detect i

on ,

se

arch,

and

a

ttention.

 Psych

ological

 Review

 

191

1 84 1-

190.

Sh eph

erd.

The Syn

o pt ic Orga

nization

 

o

the

Brain, 2n d Ed

. N

ew

Y

ork :

Oxfo

rd  Unive

rs i ty Pr

ess ,

1

979.

Shiffr

in, R. M .

, Sc

hneider, W .

C o

ntrolled and

 autom

atic hum

an info

rmation

 process

in g:

II .

 

Percept

ual learning

,

automat i

c attend

in g, and a 

general th e

ory. Psycho

lo gical R

eview,

1977

,

84

, 127 -19

0 .

Si

mon, H .

A .

T he arc

hitectur

e

of

compl

exity. Proce

ed ings

o th e

 

Am

erican

Philoso

ph ical

Soc

iety,

196

2,

26,

4

67-482 .

W

J C h

. 3. H C

A : Prelim

in ary

draft o

f 04

 

A

ugust

87   17

:44. Limited

 distrib

ution

D o

not

q

uote