Life
Games and Statistical Models
(evolution equations/pattern/order/stochastic elements/competition)
M. DRESDEN AND D. WONG
Institute for Theoretical Physics,StateUniversity of New York at Stony Brook, StonyBrook,NewYork, 11794 CommunicatedbyMark Kac, December 19, 1974
ABSTRACT Aset.ofequations isobtained, which de-scribes the rules of a class of games (life games). These games simulate the processes ofgrowth, death, survival, and competition. The equations are nonlinear difference equations, where the degree of nonlinearity is directly relatedto the number of interacting neighbors. The time evolution andthe development ofgeometricpatterns can be studied starting from these equations. Extensions and generalizations, such as the introduction of stochastic elements, can easily be accommodated in the formalism. Some significant unsolved problems are noted.
1. Background, the connection between life games and statisticalmechanicalmodels
Eversince vonNeumann's basic studies (1) there has been a continued interest in games simulating births, growth, and death, as models for life processes. An especially interesting game ofthis type was invented by J. H. Conway (2, 3); he called it "life" because of its suggestive similarities to the time evolutionof asocietyof-or organized group of-living organisms. In this game the universe considered is
two-dimensional,
divided into a large (infinite) number- of cells.Each cell can exist in two states, occupied or unoccupied, roughly corresponding to life and death. The rules of the gameregulate justhowdeaths andbirthsinagiven cell shall occur in terms of theoccupancy of theneighboringcells. The stateofthe system atany one time t is,therefore, determined by thesequence {
ni(t)
}, where i labels the cells, andpi
can be either 1 or 0 (for occupied or empty). The rules are fullydeterministic;
a specification ofthe state at time t gives anunambiguous, precise, determination
of
the state at time(t+ 1).
MorerecentlyEigen (4)in apaper
on,
theoriginofbiological
information referred to the "life" game as an interesting exampleof adeterministic scheme, which, although basedon verysimple rules,yetgivesriseto anextraordinary varietyof
configurations
andtypes ofbehavior.In fact, the behavior intime and thegeometric patterns
of
thegameare soinvolved,
thatnogeneral way tosurvey them seems to
exist;
the bulk of the surprising and at times amazing results have been obtainedfrom computer studies. As stressedagain by Eigen(4),
the furtherutility
ofthe "life" game instudying
actual (biological)evolutionary
processes is limited because of the absence of stochastic elements in the scheme. In actualbiological
processes afluctuating
environment isalways
present. Given the incompletenessofthepresent
description,
it would appear of great interest to obtain a formal or an
analytical
description
of this and related games, tosupple-ment the computer information.
General
resultsareusually
more easily obtained from formal mathematical consider-ations than from computer programs. There are
interesting
andintricategeometrical questions which needstudy, before
a
thorough
understandingcanbe obtained. Finally,stability
questions are of great significance; these again require ana-lytical tools. By combining thesemethods one might obtain new insights in the detailed operation and development of lifeprocesses.
The main point of this paperisthe observation that the formalism and the methods developed in connection with problems in nonequilibrium statistical mechanics (5-7), are particularly well suited for an analytic description of the "life" games. Moreimportant than that, analyticformulation provides (at least inprinciple) the possibility of a statistical treatment, the possibility of the systematic introduction of stochastic elements, and the possibility of substantial ex-tensionsand alterations of these games. It is notreally sur-prising that results of thestudy ofnonequilibriumbehaviorof models would be of relevance in "life" games. The "rules" ofa game are just a particular version of a discrete dynamics. It must therefore be possible to transcribe the game rules asmechanical equations of motion, yielding a direct relation between the life games and statistical models.
In the statistical problems the time dependence of various physical quantities (such as the total number of particles in a given state) are of great interest. The same type quantity is also of great importance in life-type games. For example,the total number of individuals at time t of the "life" game
N(t) =
E
qv(t)
p [1]
governs the indefinite growth, extinction, or oscillatory pattern of thesystem. Fromcomputer studies all three types of behavior are known to occur for appropriate initial con-figuration; however, computer studies by themselves cannot yield a connection between the nature ofN(t) and thenature ofthe initial state.
Inthenextsection (section 2) the methods of refs.5-7 are used to obtain the equations of motion for a "one-dimen-sional" version of the lifegame. This is done to illustrate the procedure andmethod;"one-dimensional life"byitself isnot particularly exciting.Thesamemethods, however, areuseful in the construction of the equations of motion for the full "Conwaylife" game. Theseequationsare oneof theprincipal results ofthispaper;theyarecontained in section 3. Once the formal structure of these equations is recognized, several generalizations suggestthemselves, which seem toyieldmore appropriate descriptions ofevolutionary processes. The most interesting generalizations result from the introduction of stochasticelements, orthe introduction ofcompetingspecies. It is not toodifficult to constructthe
appropriate
equations,LifeGames and Statistical Models 957 buttheirmathematical discussionisfar fromstraightforward.
Some comments together with some questions raised by the equations are contained in section 4. One of the interesting insights to emerge from this discussion is the distinction between the model (or mechanical) rules, which presumably express physical or chemical laws; the evolutionary pattern, which is a mathematical consequence of the dynamical equations; andthe rulesofreward,which define the notion of success or survival in competing systems. It is conjectured that the essence of the notion of life is to be found in the abilitytoalter the rules of reward.
2. Formalismof the one-dimensional "life"game
Inthe "life"game as originally formulated,a
given
cellin a two-dimensional grid has 8 neighbors. The survival, death, andbirthrulesarephrased
intermsofthenumber ofneighbors
of a cell. If acell isoccupiedattimet,andhas0,or 1neighbors it willbeunoccupied attimet + 1 (the cell dies from isola-tion). If a cell is occupied attimetand has 4, 5, 6, 7, and 8 neighbors, itagain will beempty attimet + 1 (the cell dies from overcrowding). If a cell is occupied and has 2 or 3 neighbors it will survive, from time t to t + 1. A cell un-occupied attime tand
having
exactly
three occupiedneigh-borswillbeoccupiedattimet+ 1. Todescribethis* introduce thevariables
...
.p(t)
= 1 if pisoccupiedattimet [2a].7.
rp(t)
= 0 if p is empty attime t [2b] Clearly, what is wanted is anequation
of motion forsp(t
+
1)
intermsofqp(t)
whichincorporates
therules of the game. To illustrate themethod,
considerfirst a one-dimen-sionalversion ofthe "life"game.t
Inthatspecial
caseeachcell p hasjusttwoneighborsp-1,p+ 1.Introduce the number ofneighbors ofp:
Vp(t) -p +
I(t)
+np_(t)
[3] Theone-dimensional "life" rulesare now:Ifp isoccupiedattime tand Vp = 0, or
vp
2, then p will be empty attime t+ 1.Ifpisoccupiedattimetand
vp
= 1,p willremainoccupied attime t + 1.Ifp isunoccupiedattimetandVP = 2,p willbeoccupiedat timet + 1.
It is notdifficulttoverify that these rulesareallcontained inthe equation:
77(t
+ 1) = 7P-1(t)70(t)
+-qp
-i(t)
77p
+i(t)+ 7p(t) 77p+i(t) -
377p
-1(t)
7ip(t)
tip +1(t)
[41Eq. 4 represents a set of coupled nonlinear difference equa-tions. Before commenting on these equations, it isfor later applicationsuseful torewritethegame rules in terms of
vp:
7tP(t
+1)
=vp(2
-Vp)f7p
+'/2vp(vp
-1)(1
-tip)
[5]Eq.5expressesthat independentof
-qp(t),
tip(t
+ 1)vanishes if VP = 0. (If there are no neighbors there can be neithersur-vival nor birth.) If
vp
= 2,qp(t
+ 1) = 1 -qp(t),
again in harmony with the rules. vp = 1 givesqp(t
+ 1) =+p(t)
one neighbor perpetuates the status quo. It is generally easier to phrase the game rules in terms of the v variables,
thendirectly in terms of the n variables. It was possible to guess Eq. 4, but it is very hard to guess higher dimensional equations.
Eq. 4 is a well-defined deterministic set of equations; it is notdifficult to see that the state at p at t isdetermined bythe initial values-qp-t(O).
p(O). *p
+t(O)accordingtoP(t)=
co(t) +EcI(t)al
+E'cij(t)aiaj
+ ...+
a~p
aP.
ap.
e +t [6] In[61
7p
-I(O)
=ap
-t, thecl(t)cij,
aretime-dependent func-tions, whichcanbe showntosatisfy
linearalgebraic
relations. Thesummationsin [6]all run overij,fromp - ttop + t.A state of the system is described by a sequence ofone's and zero's;forexample,
state = 1 0 0 1 1 1.. .1 [7] Asnoted insection1, thegeometricalpatternswhichdevelop inthisgame are notterribly interesting. Onecanshow either from the rules or from [4] that an initially finite configu-ration remains finiteor decays, butnevergrows indefinitely. Forexample, thepatterns
11 0 0 1 1 0 0 0 11 0 0 11 remainconstant intime while
1 1 1...1 decaysin twosteps.
Aconfigurationsuch as
[8a
]
[8b
I
1 1 0 1 0 0 1 1010[8c]
inwhichtwoemptyplaces(two zero's)occurconsecutively isa separated
configuration.
It is easy to see that an initiallyseparated
configuration
remains separated.Thus,eventhisverytrivialcaseexhibitsavariety of possible
behaviors.
It is interesting to note that it is very easy toaccommodate two
competing
species within this formalism. Assumethateachlocationpcouldbe occupiedby two differ-ent types of entities. Denote their numbers bytp(t)
andtip'(t).
Assume that all the rules are asbefore,
except ifeither typehasanyneighbor of the othertype, it will die, at the nextstep. Inthat casethesystemisdescribed
by
tiP(t
+ 1) = 1/2(P'p -2) (v'P - 1)[vp(2
- Pp)np+ '/2V (P-1)((- p)] [9a]
r7p'T
+1)
=1/2(Vpp-2)(vp
-1)
[P'(2
-'p)
77'p
+ 1/2VIp(Vp-1) (1 -t'p)] [9b]
The definition of v'p is of course
v'p
=(n'p
+ 1 +i'p
- 1).Eq. 9
describes
a competing coupled system, whichis,how-ever, still
completely deterministic.
It is also possible to introduce probability notions by introducing ameasure on the space ofsequences, or
by
de-fining.(...
7t,,t),
theprobability
that the system is de-scribedby
aconfiguration
1...i,,,
andderiving
anequation
for Walong the linesofrefs. 6 or7. This will becarried out in a later publication, but even these brief remarks should
*These variables are of the same type as those introduced in
refs. 6 and 7.
tJust the-one dimensional case is treated in the remainder of thissection.
958 Biophysics: Dresden and Wong
show that thesegeneralizationsare notdifficult to formulate within this framework.
3. Formalism of the full
"lifH"
gameToconstruct the equations of motion for the full "life" game, it isbestto use thesecond method used in the last section, i.e., the formulationof the game rules, interms of the number of nearestneighbors Pp.In the general case
v,,
isdefined by8 ip= , p+ i
i-= [10]
of motionin other cases. In the general case when the game rules are death survival death birth if v =0Q
1, 2,.
..a ifv =a+1...b
ifv =b+1...n if v =cc+1...c+A
theequationhas the samestructureas
[11
]:
7p(t
+
1)
=N1(v)qp(t)P(l)(vp)
+
N2(P)[1
-P(2(Vp)I(
Thesumintended in [10] is over the nearest neighbors of p. Using [10] it can be shown that the full game rules are expressed bythefollowingequationsof motion:
[Ila] vp '= (E7p+i)
-,i
7p(t
+
1)
=XlPl(vp)
+
X2-qpP2(vp)
Here
Xi
and X2arenumbers:XI=1
1Xl= \2=
-720 1440
[lb]
[12]
In
[11]
Pi
andP2 arepolynomialsinvp:
Pi(v)
=(8
-v)(7
-v)(6
-v)(5
-v)(4
-v)(2
-v)(1 -V)v
8 3(v-
V)
= _n=O 3-v[13a]
P2(v) = (8 - v)(7 -v)(6-v)(5~~~~~8
-v)(4 - v)(3-v)(1 -v _ P)
[13b]
-2 - vThesesomewhat curiouslooking polynomialsare constructed toexpress the game rule
conveniently.
ForexamplePi(v)
is alwayszero(i.e.,
forallv)
unlessv = 3. In that caseP1(3) = 720. Clearly P2(3) = 0. Thus[ii]
asserts thatv=
3--n(t+ 1) = 1[14]
This is
precisely
what thegameruledemands. For ifphas threeneighbors
and isunoccupied,
the model rulesdemand thatatthenextstep, pwill beoccupied.
Ifphasthree
neigh-borsand isoccupied,
it willremain sofor thenextstep,again
yielding
q (t+ 1) 1.Thisreasoningverified
[11
]
forv = 3; it should besimilarly
verified
for all other v. This is quite astraightforward
pro-cedure;
forexample,
Pl(v),
P2(p)
both vanish when v=8,
7,
6,5,4,1,0; then
according
to[11
]7p(t
+1)
=0,
independent
of
lip(t).
This expresses thegame rulethatforthoseparticular
numbers ofneighbors
anoccupied
cell becomesemptyeither
from
overcrowding
orisolation.
The verification of[11]
forv = 2 followsthat same
pattern,
andis omitted.Remark 1: It should
be
clear fromthis
result,
that onecan construct
analogous
equations
foranynumber
ofnearestneighbors.
This number n has nospecial
geometrical
signifi-cance; it is
only
related to thedimensionality
of the space through theparticular
lattice structureconsidered;
it could be8 as intheplane
lattice
consideredby
Conway,
and4inaspatial
(tetrahedral)
arrangement.
It is astraightforward
extension of the arguments
just
given
toconstructequations
-
lp)
[15]
Now thepolynomials havethe form:P(l)(v)
= (v -1) ...(v-a)(v
- b). [v- (b +1)]
*(v-v
[16a]
P(2) y=
'(Y
- 1) ( -a)[v-(a+ 1)]*. * [-(c+1) ][v-(c+A+1 )
*
(v-b)
... v)][16b]
The
N(v)
arenormalizationsasbefore; determined by NI(v) =[P(l)(0)>1
fora + 1 < v < b - 1 [17a]N2(v)
=[P(2)(v)]1
for c < v < c + A[17b]
In [15] the independent variable
Pp
is thesum [Ila] extended from 1 ...n. The normalization constant depends on thenumber
ofnearestneighbors.
Remark 2: Even though the
polynomials
in[13]
appear tobe ofdegree
8 it should be remembered that 7 2 =So
that the 8thpowerof
(EZ
Fp +{)
containsasmostcomplicated
terma
product
of the qvariables of the8neighbors
ofp.TheEq.1lbthencontainsa
(symmetric)
sumofproducts
ofthe
vq
variables
of theneighbors
ofp,with certainnumerical
constants. Thus
[lib]
isthe precise
counterpart ofEq.
4for theone-dimensional
(better
two-neighborl)
"life"game.Remark 3: It is
worth
stressing
that thetechniques
de-veloped
inthis
section allow still moregeneral
systemstobe describedinmuch
thesame manner.Asanexample,
considera system where cells can beoccupied
in two states(male,
female,spin
up,spin
down, charge
+,charge
-).
Theappropriate
variables arenow[18]
[19]
-1
7p= 4 0t+1
pp =E7ip
+, P = S1P+ iClearly, vp
measuresthetotalnumber
ofneighbors,
irrespec-tive ofsex
(or charge)
whileO1,
measures the excess sex(or
charge).
Further,
forsurvival
it isrequired
thata< vp < b
and
-a XOp
< +a survival[20a]
c
Pvp
. c + / and -\ .OP
+,
I birth[20b]
p<
a,orvp > bI@PI
> a death[20c]
The sexofa birthshall be(-0,,),
in case0p
= 0 it shall be female(negative).
Finally
anatural
lifetime of eachstateisintroduced,
calledT, whichmeans
that
astateqpcannotpersist
longer
thanT. Proc.Nat.Acad.
Sci. USA 72(1975)
LifeGames andStatistical Models 959 Thus, acondition for survival ist, < T;anadditionalcause
of death islongevity:
tp> T
Theequationsforthissystem areobtained in thesameway andthey read:
,2 +
'(tp
+1) =7p N(
p)P(
p)M(II(op)Q(')(OP)
X
{[1
-0(tp
-m)][e(t.
-(m
-1)T]}
- lion 0P+
E-
[1 -(vip(m))2]N(2)P(2)M(2)Q(2)
[21]ois the usualstepfunction. Here the Pand Qarepolynomials
PM(l)(p)=
p(vp
- 1)...(vp
-a)(v.
-b)(vp
- b- 1) ...(,p- n) [22a]pM(,V(P)
jVP.. (,VP - C)(,P - c- (,VP - c- A - 1) ...(,P-n) [22b]Q(1)(Op)
=(0p2
-n2)
(02-
(n-1)2)
...(@p2
-Ca)
[22c]
Q(2)(0P)
(0P2
n2)(Op2-
(n 11)2)...(0P2
- 2) [22d]TheN and M are again normalization constants
NWl)PM = 1 [23a]
N(2,p(2)
= 1 [23b]M(1)Q(l)=
1 [23c]M(2)Q(2)
= 1[23d]
This system, still deterministic, has therefore two new features, the finite lifetime, and a "charge" limitation on life anddeath.Apreliminarycomputercalculationindicates that the total number increases sharply in the average and ap-proaches infinity, but there are strong oscillations super-imposedonthissmooth behavior.
4. Comments, unsettled questions
It would be misleading toimply that with the construction of the variousequations in section3 a great deal of under-standing has been gained aboutthelife-type games. There are
always
certain advantages associated with analyticformula-tions,
butonlyifit would be possibletodeduceresultsfromthe equations, which couldnot sodirectly be obtained from a computer
study,
wouldgenuine progress have been made. It may, therefore, be useful to raise a number of specific questions, which ought to bestudied in general for the equa-tions which have been setupt.
The real utility of the equa-tions will depend on the effectiveness with which they can handlethese questions.(a)
Oneseries ofquestions refers to the time behavior of the totalnumber ofindividuals N(t). This can grow to infinity in somecases; in onedimensionit eitherdecays, or is constant. One knows from computer studiesthatfor8neighbors varioustypesof behaviorcanoccur.But it should be
possible
froman examination oftheequations
togetasystematic
answer.(b)Inthe computerstudies of theConway "life"game very interesting and
special
geometric
configurations
arise as finalstablestates. Itwould beinteresting
tofirstseeandthen derive, whether among these states certainpatterns
occurpreferentially. One could further
study
whethersuchevolved figures orgeometrical
arrangements possess a greater ordifferentdegree of
symmetry
than theinitial
configurations.
(c) Itwould be
interesting
andimportant
tohavea system-atic study of the effect of random elements. Asexplained
before, it isnotdifficult toincorporate
this in theequations,
butcertain importantquestions
cannot(or
donotappearto)
be answered so
easily.
Suppose that without random ele-ments, certaingeometric
patternsdevelop
preferentially,
as forexample, the benzenering hexagon
intheConway
model§.
The
question
thenis,
which one of these patterns is most(or least) affected
by
theintroduction of randomelements,
ormistakesormutations. This
question
suggestsalarger
and moreambitious setofproblems;
theappropriate
definition of stability andespecially
of structuralstability
in discrete dynamicalsystems.(d) Another importantissueishowasystemwould
respond
tothe imposition ofanoverall constraint. Avery
simple
ex-ample would be a fixed maximum numberNo
ofcells,
or a fixedmaximumaccessiblevolume,
inatypical
"life"game.It isconceivablethatundersuchcircumstances,theevolutionary
patternfor
Ninitial
<N(N,
issomecriticalnumber,
N,
<No)
would be quite different from that for
Winitial
< N(however
N
<No).
(e) A very trivial
example
waspresented
of thelife-type
game
equations
with twointeracting
species.
This seems animportant extension to consider
seriously.
The interaction between the two species(or players
orsexes)
could be pre-scribedstochastically
or in a deterministic manner. It now also becomespossible
toprescribe
"scoring rules,"
or"suc-cess"
or rules of reward if these twospecies
are viewed ascompetitors. One could defineassuccesslongevity, numerical superiority, or survival.
Alternately,
the attainment of aparticular
geometrical
pattern could be defined as success.It would appear that a truelife-like feature will be added if the rules of reward could be altered while maintaining the physical (game-like) rules of operation. It isalsoconceivable thatastage isreached intheoperationorintheevolutionary process wheretherules of reward themselvesbecome
ambigu-ous, sothatabifurcation processmightoccur.
It
would
be very worthwhile if examples, no matter howcontrived, exhibiting this behavior could actually be con-structed. Some efforts inthis directionwill be reported in a
subsequent publication.
Some of the
questions
raised hereneed to be answered (or shown tobeirrelevant)iffurtherunderstandingof the nature of the evolutionary processes is to be gained. The equations setup in this papershould be auseful starting point toward thatgoal.More detailed calculations, especially aboutthe statistical features of the systems, will with appropriate luck be pre-sentedindue course.
§Instudies made on games with 5 nearest neighbors, the square appears to be overwhelmingly prevalent among stable final
states.
tItmightwellbe wise to first study these questions for the case of3or 4neighbors, rather than the full"life"game.
This work was supported in part by Grant no. P4-P2656-00 from theNational Science Foundation.
1. von Neumann, A. J. & Morgenstern, 0. (1962) Theory of Games and Economic Behavior (Princeton Univ. Press, Princeton,N.J.).
2. Gardner,M.(1970)Sci. Amer.223(Oct.),120-123. 3. Gardner,M.(1971) Sci.Amer.224(Feb.),112-117.
4.
Eigen,
M.(1973)
in ThePhysicist's
Conception of Nature,ed. Mehra, K. (Reidel Publishing, Dordrecht,
Holland),
pp. 594-635.5. Kac,M.(1956)Bull.RoyalSoc.Belgium 42,356-361. 6. Dresden,M. (1962)inStudies inStatisticalMechanics. I,eds.
deBoer,J. &Uhlenbeck, G. E. (NorthHolland Publishing Co., Amsterdam),pp.303-350.