From Cognition in Animals to Cognition in Superorganisms

ghostslimAI and Robotics

Feb 23, 2014 (3 years and 1 month ago)

53 views

The Cognit
ive Animal
--
Taylor, p
age 1
From Cogni
tion in Ani
mals to Cog
nition in S
uperorganis
ms
Charles E
. Taylor

R
esearch Que
stions
My re
search in c
ognition ha
s not been
directed to
ward animal
cognition,

per se
. Rather,
I have
been
intrigued
by the appa
rent inabil
ity of mode
rn science,
as I under
stand it, t
o explain
s
ubjective e
xperience.
For this di
scussion I
will refer
to it as "t
he problem
of mind." W
hile not th
e
same, the
re is broad
overlap be
tween this
problem and
those of a
nimal cogni
tion.
My res
earch on th
e problem o
f mind has
been motiva
ted by beli
efs, in no
way origina
l with
me,
that:

Far from c
onsciousnes
s being exp
lained by a
nyone, I do
n’t think w
e even know
what an
ad
equate expl
anation wou
ld look lik
e, an opini
on articula
ted especia
lly well by
Nagel
(198
6);

Probably s
ome animals
experience
subjective
feelings i
n more or l
ess the sam
e way that
we do, thou
gh there is
a gradatio
n;

At present
we don’t k
now how to
objectively
identify w
hether or n
ot others,
animal or
h
uman, exper
ience subje
ctive feeli
ngs or what
those feel
ings are li
ke -- more
or less for

reasons di
scussed by
Erwin Schrö
dinger (196
7) as "obje
ctivation";

The study
of Artifici
al Intellig
ence (AI) a
nd robotics
have contr
ibuted a gr
eat deal to
our
unders
tanding of
the problem
, but have
certainly n
ot solved o
r "explaine
d" it.
But i
t did seem
to me that:
The Cognit
ive Animal
--
Taylor, p
age 2

While we a
re not now
able to con
struct an a
rtificial s
ystem that
can obvious
ly experien
ce
subjecti
ve feeling,
we may sti
ll be able
to
evolve
systems th
at do, usin
g technique
s that
have
come to be
known as "
evolutionar
y computati
on" (Mitche
ll and Tayl
or 1999).
In
recent yea
rs, largely
as a resul
t of work b
y Rodney Br
ooks and hi
s students,
much of wh
ich is
coll
ected in Br
ooks (1999)
, I have co
me to belie
ve further,
that

For purpos
es of under
standing co
gnition, es
pecially an
imal cognit
ion, and pr
obably
subj
ective expe
rience, org
anisms are
best viewed
as collect
ions of sen
sors, effec
tors and
pr
ocessors of
limited ab
ilities sur
rounding --
and studde
d throughou
t -- the or
ganism,
com
municating
primarily w
ith other s
ensors, eff
ectors or p
rocesses th
at are most
ly
nearby a
nd mostly w
ith limited
bandwidth,
functionin
g together
as an ensem
ble.
This me
ans that ou
r cognition
and subjec
tive experi
ence are to
be underst
ood as the
collective
behavior of
these limi
ted agents,
acting as
one. If so,
then one o
f the most
important
c
hallenges f
acing us, b
ut one we c
an probably
address in
concrete t
erms, will
be to under
stand
how s
uch an ense
mble will f
unction as
some sort o
f "superorg
anism".
In t
he remainde
r of this e
ssay I will
attempt to
explain ho
w and why I
arrived at
this opini
on,
and at
the end, br
iefly addre
ss some of
the promise
s and limit
ations it o
ffers.
Past
Research
Ini
tially I ap
proached th
e problem o
f mind by l
ooking at a
nimals that
could reas
onably be
a
ssumed to e
xperience r
udimentary
subjectivit
y. We used
mutant stra
ins of
Drosophila

melanogast
er
, explorin
g how learn
ing or memo
ry-retentio
n mutations
affected h
abitat sele
ction
(Tayl
or 1987). T
hese were c
hosen becau
se much is
known about
their biol
ogy and bec
ause they
The Cognit
ive Animal
--
Taylor, p
age 3
are easy t
o grow and
evolve. I s
oon abandon
ed this bec
ause it see
med difficu
lt for me t
o relate to

their sens
ory experie
nces; they
were settin
g the bound
ary conditi
ons on this
, not I. It
was too
di
fficult to
broach the
subjective-
objective p
roblem ment
ioned above
.
Artificial

systems, h
owever, can
be made mo
re or less
complex, as
desired. A
t that time

there was
enthusiasm
that the "t
raditional"
approach t
o artificia
l intellige
nce would p
rovide
insi
ghts to the
problems o
f cognition
. By "tradi
tional" I m
ean the app
roach exemp
lified by N
ewell
and S
imon (1972)
in their G
eneral Prob
lem Solver
(GPS) progr
am, for exa
mple. Very
simply, a
c
ognitive sy
stem is vie
wed as havi
ng a snapsh
ot of a sta
te space en
coded by sy
mbols; ther
e are
rules
by which t
hese symbol
s may be ma
nipulated;
and cogniti
on consists
of searchi
ng through
a
sequence
of manipula
tions of th
ese symbols
so that th
ey will com
e to the de
sired confi
guration of

state spac
e in a suit
ably econom
ical fashio
n. Accordin
g to this v
iew, proble
m solving b
y animals
c
onsists of
obtaining i
nformation
about their
environmen
t (describa
ble by pred
icate calcu
lus) and
th
en manipula
ting this r
epresentati
on to obtai
n a desired
state of a
ffairs, aga
in describa
ble by
pred
icate calcu
lus. The an
imal would
then follow
that seque
nce. Cognit
ion is the
deliberatio
n
about whi
ch chain of
actions is
most appro
priate.
Ther
e is a scho
ol of thoug
ht in evolu
tionary bio
logy that v
iews the en
vironment a
s posing
pr
oblems that
animals ne
ed to solve
and that e
volution gu
ides them t
oward doing
so. Pursui
ng
this app
roach, I de
veloped a c
omputer sys
tem where o
bjects inhe
rited certa
in proposit
ions at the

time they
were create
d. Other pr
opositions
could be ob
tained from
the enviro
nment. The
objects
wou
ld then act
on these p
ropositions
using the
rules of sy
mbolic logi
c to see if
they could
solve
prob
lems posed
by the envi
ronment. Th
ose who did
so could r
eproduce, t
hose who we
re unable
t
o arrive at
those solu
tions peris
hed without
leaving of
fspring. Th
e system wa
s written i
n Lisp and
worked only
modestly w
ell, so I d
id not publ
ish it. Sin
ce that tim
e others, e
specially J
ohn Koza
The Cognit
ive Animal
--
Taylor, p
age 4
(see Fogel
1998; Mitc
hell and Ta
ylor 1999),
have devel
oped powerf
ul Lisp-bas
ed systems
for
evoluti
onary compu
tation term
ed "Genetic
Programmin
g." Koza et
al. (1992)
were able
to obtain
s
ome insight
s about (po
ssible) rea
soning by a
nolis lizar
ds while th
ey chose to
pursue ins
ects or
not
while fora
ging. Audre
y Cramer, w
ho started
her researc
h in my lab
oratory as
a graduate
student, su
bsequently
explored ho
w similar m
odels of co
gnition mig
ht explain
how vervet
monkeys cho
se a route
in their se
arch for fo
od (Cramer
and Gallist
el 1997).
Ab
out this ti
me I began
working wit
h David Jef
ferson on a
system he
had develop
ed that
emb
odies what
I feel char
acterizes t
he nature o
f organisms
and life.
We have art
iculated th
is in
sever
al papers,
summarized
in Taylor a
nd Jefferso
n (1995). M
ost importa
nt is the v
iew that li
fe is
a pro
perty of
processes
, rather t
han of
objects
. Living b
eings are s
elf-organiz
ing process
es (or
even
an ensembl
e of proces
ses) that i
nteract wit
h the world
around the
m, taking i
n some thin
gs
and expe
lling other
s, possibly
reproducin
g other suc
h processes
, and event
ually "dyin
g", in that

they are n
o longer ca
pable of se
lf-perpetua
tion. Senti
ments about
process
vs.
object tha
t
influence
d my thinki
ng about th
is include
Whitehead (
1926) and B
irch and Co
bb (1981),
though I
ha
ve been una
ble to reco
ncile certa
in signific
ant differe
nces betwee
n us. In an
y event, it
seems
like
ly to me th
at mind as
we know it
can be poss
essed only
by such "li
ving" proce
sses or obj
ects
embody
ing such pr
ocesses. If
so, then m
y prior wor
k with prob
lem-solving
had fundam
ental
short
comings.
Rob
Collins, a
student of
Jefferson’
s, used Con
nection Mac
hines -- co
mputers wit
h tens of
t
housands of
processors
, each with
its own me
mory and pr
ocesses, to
evolve pop
ulations of

processes
that seemed
life-like,
in that th
ey could re
produce, mo
ve, die and
evolve as
they
naviga
ted a virtu
al world (J
efferson et
al. 1992).
Such syste
ms could al
so learn, n
ot simply e
volve,
a co
mbination t
hat has gre
at adaptive
potential
(Belew and
Mitchell 19
96).
The Cognit
ive Animal
--
Taylor, p
age 5
About this
time Jeffe
rson and I
became acqu
ainted with
the theori
es of Rodne
y Brooks, a
n
MIT expon
ent of "Nou
velle AI".
In a series
of papers
with titles
like "Inte
lligence wi
thout reaso
n"
and "Ele
phants don’
t play ches
s" (collect
ed in Brook
s 1999) he
argued that
intelligen
t systems
n
eed to be:
(1) embodie
d, not just
simulation
s; (2) situ
ated in the
real world
, not in si
mply a
"dem
o" setting;
and that (
3) a subsum
ption archi
tecture is
the best wa
y to contro
l such syst
ems.
Brooks
’ views wer
e heavily i
nfluenced b
y certain e
thologists,
especially
von Uexkül
l (1921). F
rom
this vi
ew "traditi
onal AI" is
incapable
of capturin
g animal in
telligence.
Among othe
r reasons,
representin
g natural e
nvironments
by traditi
onal AI mod
els in a wa
y that is s
ufficiently
rich for
i
ntelligent
behavior ne
cessitates
an untenabl
e explosion
of logical
manipulati
ons. He sug
gested
that
the best w
ay for an a
nimal to re
present its
environmen
t is to use
the enviro
nment to
re
present its
elf. To tes
t such theo
ries one ca
nnot use di
sembodied s
ymbol manip
ulations, b
ut
generall
y need to u
se robots i
n rich envi
ronments. T
hese robots
have typic
ally consis
ted of larg
e
numbers o
f distribut
ed sensors,
effectors
and process
ors, each w
ith modest
abilities.
None of
the
se processo
rs could be
said to ch
aracterize
the environ
ment in any
meaningful
way by its
elf,
yet co
llectively
these syste
ms are capa
ble of comp
lex coordin
ated action
s like clim
bing up sta
irs
or coll
ecting empt
y soda cans
from desk
tops. Furth
er, watchin
g robots pl
ay tag or "
attempt" to

fetch an o
bject comes
much close
r to meetin
g the crite
ria that ar
e now used
to infer su
bjective
fe
eling, desc
ribed in th
e introduct
ion, than d
o computers
which simp
ly manipula
te symbols.

Hence we t
hought it w
orthwhile t
o explore w
hether evol
ving robots
might cont
ribute to s
tudying
the
problem of
mind.
At th
at time few
robots wer
e commercia
lly availab
le and supp
orted. So t
wo students
in my
lab,
Kourosh Na
fisi and Or
azio Miglin
o, construc
ted a small
, fairly ro
bust robot
out of Lego

blocks us
ing a small
controller
board we o
btained fro
m the MIT A
I lab (Migl
ino et al.
1994). It
The Cognit
ive Animal
--
Taylor, p
age 6
resembled
an early ve
rsion of th
e Lego Tec
hnics kits
that are no
w sold in t
oy stores.
It was
cont
rolled by n
eural nets,
and popula
tions of su
ch nets cou
ld evolve b
y a variant
of genetic

algorithms
(see Mitch
ell and Tay
lor 1999).
Miglino’s r
obot evolve
d fairly co
mplex neura
l networks
and associa
ted behavio
r to circum
navigate a
simple envi
ronment. Ev
olution of
these robot
s was
a hyb
rid between
simulated
neural nets
, similar t
o those evo
lved by Col
lins above,
and real
m
easurements
of occasio
nal physica
lly realize
d ones. We
were unable
to physica
lly test
all
of
the rob
ots that ne
eded to be
tested each
generation
, because t
he robot wa
s just plas
tic and was

not suffic
iently robu
st to withs
tand hours
of continuo
us use, nor
was it pos
sible to pr
ovide power

to them fo
r fully aut
omatic test
ing. To my
knowledge t
hese were t
he first hy
brid simula
ted/actual
robots that
were evolv
ed by evolu
tionary com
putation. S
ince that t
ime, very r
obust robot
s have
been
developed
and ways fo
und to prov
ide constan
t power (No
lfi and Flo
reano 2001)
. One of
th
ese, the Kh
epera robot
, has becom
e the workh
orse of "ev
olutionary
robotics",
though even

these are
still very
simple and
require an
environment
so artific
ial that th
ey can be r
egarded onl
y
as embodi
ed, but not
situated i
n the real
world (Broo
ks, persona
l communica
tion).
As we
attempted
to reconcil
e these stu
dies with t
hose of Bro
oks it beca
me evident
that a
crit
ical featur
e of intell
igent behav
ior by real
organisms
is the abil
ity for lar
ge numbers
of sensors,

processors
and actuat
ors to func
tion togeth
er as a uni
fied
ensemble
. The robo
t is surrou
nded
and st
udded with
many such e
lements, co
mmunicating
together,
none of the
m having a
complete
mo
del of the
environment
-- certain
ly none tha
t could be
naturally r
epresented
by proposit
ional
calcu
lus as the
earlier AI
models requ
ired -- but
each of th
em having s
ome ability
to digest
and
process
the inform
ation they
obtain loca
lly and fro
m nearby se
nsors and p
rocessors.
The
organis
m’s cogniti
on seems al
most a
collective
consciousn
ess
, derived
from the in
teraction o
f all
these
parts func
tioning as
a "superorg
anism."
The Cognit
ive Animal
--
Taylor, p
age 7
Current Re
search
If we
are to evo
lve such a
"collective
individual
" we will n
eed to pay
much more a
ttention to
just
what
is abstract
ed from the
environmen
t and commu
nicated to
other parts
of the org
anism. A
co
llection of
sensors an
d processor
s indiscrim
inately pou
ring their
electrons i
nto a bus i
s not
likel
y to achiev
e our goal.
Instead, e
ach sensor
and process
or needs to

compress
their
expe
rience that
learning o
r evolution
has judged
to be impo
rtant and t
ransmit onl
y that. It
is
especial
ly challeng
ing to make
the experi
ences of di
fferent sen
sors and pr
ocessors mu
tually
inte
lligible. W
ith a compu
tational li
nguist, Edw
ard Stabler
, and stude
nt in my la
b, Tracy Te
al, I
began
to explore
how such e
xtraction a
nd compress
ion might b
e achieved.
Our first
work has
be
en exclusiv
ely with sy
mbolic syst
ems (Teal e
t al. 1999;
Teal and T
aylor 2000)
. We are no
w
exploring
how to bin
d such symb
ols with ac
tual experi
ences. Rela
ted to this
, several y
ears ago I
cooperated
with Takaya
Arita on a
study wher
e populatio
ns of neura
l nets lear
ned to acqu
ire a
commo
n lexicon,
even when t
hey were no
t all able
to observe
the same ob
ject (Arita
and Taylor

1996).
If t
his approac
h is more o
r less corr
ect, then t
he challeng
e ahead is
to learn ho
w
experienc
es by distr
ibuted syst
ems in the
real world
are able to
combine an
d adapt to
construct a

unified wh
ole -- poss
ibly to the
extent of
making a "w
hole iguana
" (Dennett
1978). This
will
neces
sarily invo
lve enginee
ring, both
mechanical
and electri
cal, and co
mputer scie
nce with
pl
enty of the
ory and rea
l-world con
structions.
Probably l
inguistics
and evoluti
onary theor
y will
prov
ide needed
insights.
The Cognit
ive Animal
--
Taylor, p
age 8
Phylogenet
ic appropri
ateness
The
overriding
question or
test of th
is approach
is one of
phylogeneti
c appropria
teness. Is
in-silico
life going
to be suff
iciently si
milar to
in-vivo
life, in t
he way that
much
in-vitro
life has
p
roven to be
? Will thes
e simpler a
nd manipula
ble constru
ctions of i
ntelligent
behavior re
ally
going
to lead to
an understa
nding of mi
nd and cogn
ition? Or w
ill they me
rely be a d
iversion, a
lbeit
one t
hat is enth
ralling and
no doubt c
ommercially
useful, bu
t still jus
t a diversi
on, from ou
r
understan
ding of the
deeper iss
ues?
Though
poorly unde
rstood, sub
jective and
physical e
vents are u
ndeniably p
art of the
same
phenom
ena. "The w
orld is giv
en to me on
ly once, no
t one exist
ing and one
perceived"

(Schröding
er 1967). P
utting asid
e the probl
em of how s
ubjective e
xperience i
s to be rec
ognized
at
all, some p
hysical sys
tems must b
e capable o
f subjectiv
e experienc
e and some
(we believe
)
are not -
- for reaso
ns we do no
t know. We
are current
ly unable t
o judge jus
t which
might
become
cap
able and wh
ich will no
t. If a sys
tem is capa
ble of beco
ming subjec
tive by inc
remental
ch
anges, and
if our crit
eria for se
lection is
appropriate
, then it s
hould be po
ssible to e
volve
syste
ms that hav
e subjectiv
e experienc
es. But wha
t if the sy
stem is sim
ply not cap
able of the

physical i
nteractions
necessary
to support
subjective
experience?
A liter of
oxygen, th
ough
unques
tionably a
physical en
tity, is ca
pable of be
coming soli
d only unde
r the most
unusual of
circumstanc
es, if at a
ll. It migh
t be that t
he silicon
chips are s
imilarly no
t capable o
f the the
p
hysical pro
cesses that
must under
lie subject
ive experie
nces. Maybe
the necess
ary interac
tions
occur
only betwe
en carbon,
hydrogen an
d oxygen. I
f that is t
he case, th
en computer
s and robot
s,
as now c
onstructed,
will be ph
ylogenetica
lly inappro
priate vehi
cles for st
udying the
problem of
mind.
The Cognit
ive Animal
--
Taylor, p
age 9
Or maybe t
hey will ha
ve a differ
ent
sort
of subject
ive experie
nce. Extrap
olating wit
h the
imagi
native and
forceful ar
guments of
von Uexküll
(1934), we
can presum
e that the
"experience
"
of such c
reations wi
ll be quite
different
from our ow
n, and prob
ably quite
varied from
creation t
o
creation.
There is e
very reason
to expect
that these
will provid
e us an eno
rmously use
ful testbed

for studyi
ng the rela
tion betwee
n cognition
and subjec
tive experi
ence -- per
haps even p
roviding
us
evidence f
or other fo
rms of cons
cious exper
ience, call
ed for by N
agel (1986)
.
In the pas
t few decad
es we have
learned a l
ot from eff
orts to cre
ate compute
rs and robo
ts
behaving
in ways th
at might se
em to requi
re subjecti
vity. One n
eed only pe
ruse the ro
bots on
dis
play at htt
p://www.and
roidworld.c
om/ to be i
mpressed by
their vari
ety and acc
omplishment
s.
Most emp
hasis has b
een on
human
subjectivi
ty, but the
re has been
progress f
rom attempt
s to
explor
e intellige
nt behavior
and subjec
tive experi
ence at oth
er levels -
- including
lizards. T
here is
no
reason to b
elieve this
progress w
ill stop, a
nd there is
plenty of
reason for
optimism. W
hether
this
category o
f model is
truly phylo
genetically
appropriat
e or not, o
nly time wi
ll tell.
Ref
erences
Arit
a, T. and T
aylor, C. (
1996). A si
mple model
for the evo
lution of c
ommunicatio
n. In
Evolutiona
ry
Programm
ing: Procee
dings of th
e Fifth Ann
ual Confere
nce of Evol
utionary Pr
ogramming
, ed. L.
F
ogel, P. An
geline, and
T. Bäck, p
p. 405-410.
Cambridge,
MA: MIT Pr
ess.
Belew,
R. K. and M
itchell, M.
(1996)
. Adaptive
Individual
s in Evolvi
ng Populati
ons: Models
and
Algori
thms
. Reading,
MA: Addiso
n Wesley.
The Cognit
ive Animal
--
Taylor, p
age 10
Birch, C.
and Cobb, J
r., J. B. (
1981)
. The Libe
ration of L
ife
. Cambridg
e: Cambridg
e Universit
y
Press.
Bro
oks, R. A.
(1999)
. Cambrian
Intelligen
ce: The Ear
ly History
of the New
AI
. Cambridg
e, MA:
MIT
Press.
Crame
r, A. E. an
d Gallistel
, C. R. (19
97). Vervet
monkeys as
travelling
salesmen
. Nature
387:
464.
Dennett, D.
C. (1978).
Why not th
e whole igu
ana?
Behavioral
and Brain
Sciences
1: 103-10
4.
Fogel, D.
B. (1998)
. Evolutio
nary Comput
ation; The
Fossil Reco
rd
. New York
, NY: IEEE
Press.
Jeffe
rson, D., C
ollins, R.,
Cooper, C.
, Dyer, M.,
Flowers, M
., Korf, R.
, Taylor, C
., and Wang
, A.
(1992)
. Evolution
as a them
in artifici
al life: Th
e genesys/t
racker syst
em. In
Artificial
Life II
, ed. C.
G
. Langton,
C. Taylor,
J. D. Farme
r, and S. R
asmussen,

pp. 549-57
8. Reading,
MA: Addiso
n
Wesley.
Ko
za, J. R.,
Rice, J. P.
, and Rough
garden, J.
(1992). Evo
lution of f
ood foragin
g strategie
s for the
C
aribbean an
olis lizard
using gene
tic program
ming
. Adaptive
Behavior
1: 47-74.
Miglino, O.
, Nafasi, K
., and Tayl
or, C. E. (
1994). Sele
ction for w
andering be
havior in a
small robo
t
.
Artifici
al Life
2: 101-11
6.
The Cognit
ive Animal
--
Taylor, p
age 11
Mitchell,
M. and Tayl
or, C. E. (
1999). Evol
utionary co
mputation:
An overview
. Annual R
eview of
Ec
ology and S
ystematics
30: 593-6
16.
Nagel, T
. (1986)
. The View
From Nowhe
re
. New York
, NY: Oxfor
d Universit
y Press.
New
ell, A. and
Simon, H.
A. (1972)
. Human Pr
oblem Solvi
ng
. Englewoo
d Cliffs, N
J: Prentice
-Hall.
Nolfi
, S. and Fl
oreano, D.
(2001)
. Evolutio
nary Roboti
cs: The Bio
logy, Intel
ligence and

Technology
of Self-Or
ganizing Ma
chines
. Cambridg
e, MA: MIT
Press.
Schrö
dinger, E.
(1967)
. What is
Life? Mind
and Matter
. Cambridg
e: Cambridg
e Universit
y Press.
Tay
lor, C. E.
(1987). Hab
itat select
ion within
species of
drosophila:
A review o
f experimen
tal
finding
s,
Evolutiona
ry Ecology
1: 389-40
0.
Taylor, C
. E. and Je
fferson, D.
R. (1995).
Artificial
life as a
tool for bi
ological in
quiry. In
Artificial

Life: An O
verview
, ed. C.G.
Langton, p
p. 1-13. Ca
mbridge, MA
: MIT Press
.
Teal, T.,
Albro, D.,
Stabler, E.
, and Taylo
r, C. (1999
). Compress
ion and ada
ptation. In

Fifth
Euro
pean Confer
ence on Art
ificial Lif
e
, pp. 709-
719. Heidel
berg: Sprin
ger-Verlag.
The Cognit
ive Animal
--
Taylor, p
age 12
Teal, T. a
nd Taylor,
C. E. (2000
). Effects
of compress
ion on lang
uage evolut
ion
. Artifici
al Life
6:
129-14
3.
Uexküll,
J. von (192
1)
. Umwelt u
nd Innenwel
t der Tiere
. Berlin.
U
exküll, J.
von (1934).
A stroll t
hrough the
worlds of a
nimals and
men. In
Instinctiv
e Behavior:

The Develo
pment of a
Modern Conc
ept
, ed. C. H
. Schiller
(1957), pp.
5-80. New
York, NY:
I
nternationa
l Universit
ies Press.
W
hitehead, A
. N. (1926)
. Science
in the Mode
rn World
. New York
, NY: Macmi
llan.