Static and Dynamical Equilibrium Properties to Categorise Generalised

Game-of-Life Related Cellular Automata

K.A. Hawick
Computer Science, Institute for Information and Mathematical Sciences,
Massey University, North Shore 102-904, Auckland, New Zealand
email: [email protected]
Tel: +64 9 414 0800 Fax: +64 9 441 8181
April 2012

ABSTRACT
Conways’s Game-of-Life (GoL) is one of a family of spa-
tial cellular automata rule sets that employ the Moore
neighbourhood on a rectilinear lattice of cells. GoL gen-
eralises to set of other automata using different neighbour-
numbers that lead to birth and survival. These models ex-
hibit different static properties and also different long term
dynamical equilibrium points. We define some quantifi-
able metrics and explore how some well known GoL vari-
ants can be categorised into groupings, categorised by their
static properties and dynamical equilibria. We also explore Figure 1: Coral and Diamoeba Automata.
effects due to differing initial crowding levels of the live
has a concise specification, but it is formulated in terms of
seed cell population, and how this affects the categories.
counts of live neighbouring cells rather than specific pat-
terns.
KEY WORDS
statistical mechanics; Game of Life; generalised rules; cel- There is a space of similarly formulated automaton rules
lular automata; complex system. [24] in the same family as GoL [22]. Like the Wolfram
automata not all the rules in the GoL family are “interest-
ing” or can be classified as complex. The Conway precise
specification turns out to be special in the rule set space
1 Introduction of the family, but there are some other related automata
Cellular Automata (CA) models [1, 13, 28] have long in the family that do have interesting properties and which
played an important role in exploring and understanding have been given names. Figure 1 shows snapshots from
of the fundamentals of complex systems [17, 33, 34]. two such GoL family members - known as “Coral” and
“diamoeba.”
Studies of many CA systems are reported in the litera-
ture and for many complex systems applications includ- There are a number of other variants of cellular automata
ing: damage spreading [21]; network analysis [38]; food including: the game of death [12], with an extra zombie
chain systems [16]; lattice gases [37]; and fluid flow [7]. state added; ratchet automata [15]; and other variants of
Two fundamental and classic CAs that provide a basis for GoL such as hyperbolic life [27]. GoL and its variants have
much other work are the Wolfram’s one dimensional au- also also been studied: on other non-square geometries [5];
tomaton [36], and Conway’s Game of Life [14]. The Wol- on Penrose tilings [25]; on hexagonal lattices [4]; and on
fram system is especially useful because of it conciseness generalised networks [9]. Work has also been reported on
and the fact that the whole automaton rule state space is three dimensional GoL variants [2, 3].
easily described as a family of bit valued cell-wise truth ta-
This present article reports work exploring some of the
bles and can be systematically explored. Two dimensional
transient and long term dynamical properties of the GoL
automata can be also be formulated using the Wolfram ap-
family of automata in two dimensions and on a square, pe-
proach [26]. Conway’s Game of Life (GoL) however, also
riodic lattice. In particular a preliminary attempt is made
to see how this GoL family can be categorised according
to some quantitative metrics that can be obtained from nu-
merical experiments. Attempts to automatically classify
automata by their behaviours have been reported [19], al- 1 2 3
though here we discuss a set of manually carried out ex-
periments. 4 5
Much work has been done on studying the coherent struc-
tures that occur in GoL and its variants [11]. It is possi- 6 7 8
ble to implant specific patterns such as gliders, glider guns
and so forth to obtain specific sequences of GoL automata
configurations. However, in this present paper we investi-
gate GoL automata systems that have been randomly ini-
Figure 2: Moore neighbourhood of 8 sites in 2-D.
tialised with an initial fraction of live cells. Providing we
simulate a large enough sample of large enough automata
systems, many different individual patterns can occur by elaborate these and use the more verbose form of
chance, will interact and the system will eventually arrive Bn1 , n2 , .../Sm1 , m2 , .... This notation specifies the
a static or dynamical equilibrium. This approach supports rules that area applied to each site exactly once at each
study of the statistical mechanical properties [31] of the time step. The notation gives a comma-separated-list of
automata systems and the identification of which automata the allowable numbers ni of live neighbours which give
have stable points [6] or not, as well as a framework to rise to Birth of a live site from a dead site before the slash,
study universality properties [32] in cellular automata the- and likewise another comma-separated list of the allowable
ory [35]. number mi of live neighbours that are necessary for a live
site to survive.
Although stochastic thermal or random effects [18, 20]
have been introduced into cellular automata to study such Algorithm 1 Synchronous automaton update cells algo-
properties, in this present work we restrict ourselves to en- rithm.
tirely deterministic automata. We avoid asynchronous au- for all i, j in (L, L) do
tomaton [8, 23, 29, 30] effects by using a two phase syn- gather Moore Neighbour-hood Mi,j
chronous update implementation of the automata reported apply rule b[i][j] ← {s[i][j], Mi,j }
here. The only randomness or ambiguity introduced is in end for
the initial random pattern of live cells seeded. for all i, j in (L, L) do
copy b[i][j] → s[i][j]
The article is structured as follows: In Section 2 we de-
end for
scribe the notation for the GoL family of automata and the
statistical metrics that can be applied to study them numer-
ically. In Section 3 we present some results for some of
Algorithm 2 CA Experimental procedure
the automata in the family. We discuss the implications in
Section 4 and summarise some conclusions and areas for declare s[L][L]; b[L][L]
further work in Section 5. for r = 0 to runs do
initialise s, st nL = pN live, rest dead
for t = 0 to te do
automaton update cells
2 Automata Notation & Metrics end for
The game of Live and its family of variant automata are for t = te to tm do
implemented using the Moore neighbourhood on a d = 2 automaton update cells
dimensional array of cells, where the number of (Moore) make measurements
Neighbouring sites NM = 8, for d = 2 as shown in Fig- end for
ure 2. We define the (square) lattice Length as L and hence end for
the number of sites N , typically N = Ld . We define normalise averages
the number of live sites NL , and so the metric fraction
fl = NL /N and similarly the number of dead sites ND , The rules thus support death both from “overcrowding”
and fraction fD = ND /N . and also from “lack of parents.” The classic Conway game
The GoL community has developed various gener- of life is thus specified by the notation B3/S2,3. We use
alised notations for a GoL-like automata [1, 24]. We this notation to specify explicitly the 26 GoL variants we
experiment with below. Although GoL automata are well
studied, for completeness and to avoid ambiguity, we give
the algorithm for implementing these e automata.
Algorithm 2 shows the experimental procedure we deploy,
based upon the automata update rule procedure specified in
Algorithm alg:synch-update. We use a two-phase update
algorithm and so each rule is applied to change the site at
time t + 1 based on its state and that of its neighbouring
sites at time t.
The following metrics are used to study the automata ex-
perimentally:

• theP1-step time correlation function C1 =

1
N i,j s(i, j, t).s(i, j, t − 1)

• the fraction of neighbours that are the same as the site

they surround. This is essentially a count of the like-
Figure 3: 1-Step Correlation and static fractional metrics.
like bonds, normalised by divining by (N.NM )
• likewise, the fraction of neighbours that are different
from the site they surround
• the number of monomers is the number of live sites
that are completely surrounded by dead sites.
• the number of dimers is the number of connected-
pairs of live sites that are each otherwise surrounded
by dead sites.

The number of arbitrary sized cluster components of live

sites is expensive to compute and we do not use it in this
study.

3 Experimental Results
The GoL family of automata were investigated numeri-
cally. The reported experiments are based on averages over Figure 4: Monomer and Dimer counts.
100 independently randomly initialised runs of a periodic
2562 cell system, equilibrated for 512 time steps and then Figure 4 shows the number of monomers and dimers of
run with measured metrics recorded and averaged over a live cells for the different GoL automata with a starting
subsequent 512 steps. initialisation fraction of p = 0.5. The curves are again
ranked and points labelled by automata models. As can be
Table 1 shows 26 members of the GoL family of automata
seen most models do not lead to large numbers of dimers or
simulated for the specified times t on a 2562 periodic lat-
monomers, but a small number do exhibit large numbers.
tice, with random initial fractions of live sites of the speci-
The grouping is split by a relatively sharp shoulder in the
fied fractions p.
curves.
Figure 3 shows some of the metrics described in Section 2
Figure 5 shows a scatter-plot of the 1-step correlation met-
applied to the family of automata. The plots are of the
ric plotted against the fraction of live sites. There are clear
1-step correlation function; the fraction of live and Va-
and obvious groupings with some annotations drawn (in
cant(=dead) cells and the fraction of neighbour links that
red, online version) indicating which automata behave sim-
have the same (live-live or dead-dead) cell on them. The
ilarly.
plots are sorted according to rank and the labelled points
indicate the automata model. Note that the plot lines natu- The plotted metrics are shown with error bars based on the
rally group onto plateaux with sharp cliffs separating them. estimated experimental standard deviations. These them-
Automata of similar properties thus group together. selves give some interesting insights into the behaviours
Rule Common t=128 t=256 t=128 t=256 t=128 t=256
(B/S) Name
format for Rule p=0.5 p=0.5 p=0.3 p=0.3 p=0.1 p=0.1

B1/S1 Gnarl

B1,3,5,7/S1,3,5,7 Replicator

B1,3,5,7/S0,2,4,6,8 Fredkin

B2/S Seeds

B2/S0 Live Free/Die

B2,3,4/S Serviettes

B3/S0,1,2,3,4,5,6,7,8 Life w/o Death

B3/S1,2,3,4 Mazectric

B3/S1,2,3,4,5 Maze

B3/S2,3 Conway

B3/S4,5,6,7,8 Coral

B3,4/S3,4 34 Life

B3,4,5/S4,5,6,7 Assimilation

B3,4,5/S5 Long Life

B3,5,6,7,8/S5,6,7,8 Diamoeba

B3,5,7/S1,3,5,8 Amoeba

B3,5,7/S2,3,8 Pseudo Life

B3,6/S1,2,5 2x2

B3,6/S2,3 HighLife

B3,6,8/S2,4,5 Move

B3,6,7,8/S2,3,56,7,8 Stains

B3,6,7,8/S3,4,6,7,8 Day & Night

B3,7,8/S2,3,5,6,7,8 Coagulations

B4,6,7,8/S2,3,4,5 Walled Cities

B4,6,7,8/S3,5,6,7,8 Vote4/5/Annl

B5,6,7,8/S4,5,6,7,8 Majority

Table 1: 2562 Automata for different rules at times 128 and 256, for p = 0.5, 0.3, 0.1.
 2
1
3a

3b

Figure 5: Scatter-plot of fCorr vs fLive, with annotated groupings in red (colour, online).

of the different automata. Models like “coagulation” and subsequently reaches a saturated state
“anneal” have the highest fluctuations - due to the slowly
but continually varying spatial pattern observed in those 3. a dynamic equilibrium of structures that fluctuate and
automata over the time frame of the measurements. gives rise to a permanently changing pattern - these
refine into a) those with large scale structures and b)
those that only exhibit short scale fluctuations
4 Discussion Table 2 was originally constructed somewhat subjectively
One long term goal of this approach is to identify the dy- after examination of the time sequences of patterns ob-
namical growth and equilibria properties of the different tained from the 26 automata. The method involved eye-
automata [10]. To this end it is useful to attempt a group- balling the models as movies trying to describe some com-
ing of classification based on temporal as well s spatial be- mon words from the observations. Somewhat amazingly,
haviour. The experimental data gathered indicates that the these however do appear to agree with the clusters found
GoL automata studied fall into one of the following cate- upon examination of Figure 5. The category 3a - “dynamic
gories: equilibrium with long and multi-scale structures” cluster
together at the top left of the scatter plot.
1. rapidly reach a single saturated state of almost all
The category 3b - “ dynamic equilibrium but with short
dead or live with at most minor fluctuations
length scale structures” cluster in a larger area at the lower
2. steady growth in a complex structural pattern that then left. The other two groups are at the top right area with
 Category Game-of-Live Family Examples - Categorised by Subjective Observation
1 Assimilation; Diamoeba; Live w/o Death; Replicator
2 Coagulations; Coral; Day+Night; Majority; Maze; Mazectric; Move; Anneal
3a 2x2; Conway; Highlife; Pseudo Life
3b 34-Life; Amoeba; Fredkin; Gnarl; Live Free or Die; Long Life; Seeds; Serviettes; Stains; Walled Cities

Table 2: Categorisation of GoL Automata for p = 0.5

some differentiation between fast and slow equilibration at

the far right and middle right respectively.
5 Conclusions
The single exception to this grouping is the “replicator” We have shown how a family of Game-of-Life-like Cellu-
automata. It exhibits the strange behaviour of remaining lar Automata can be formulated in terms of a common and
in fluctuation for a long time then suddenly saturating to a extensible notation and systematically studied. We have
single dead state, and this may be due to finite size effects shown that some simple metrics help make preliminary at-
of the simulations. tempts to categorise this family of automata into groups.

In Figure 4 the monomers and dimer curves are less ob- We have postulated four categories – originally conceived
vious discriminators by themselves. The category 3a au- in terms of a mix of temporal and spatial structural ob-
tomata do however group closely on the mid or upper servations. These map quite closely to the groupings that
shoulder of the monomer and dimer curves, respectively. emerge from a scatter plot of the long term averaged values
of the fraction of live cells and the 1-step time correlation
Similarly, in Figure 3 the single metric averages do not pro- function.
vide good discriminators individually, although there are
obvious groupings of the 3a automata appearing together, There are interesting features in all the GoL automata stud-
and the category 1 also tend to be close together. ied - this is no doubt why the community (See http:
//www.conwaylife.com/wiki) has troubled to give
We might expect that a caveat of this analysis is that when these models specific names. However, perhaps the “most
the system is not sufficiently “thermodynamically large interesting” ones are those that seem to exhibit the high-
enough” then: est degree of complexity and not-coincidentally have been
given the names with the word “life” in their names. These
• the pattern might otherwise continue to grow if it did category 3a automata stabilise to a dynamic equilibria of a
not reach the periodic boundaries relatively low fraction of live cells, with a high degree of
time correlation between their states.

• the particular set of structural patterns that might lead There is scope for further work in examining time-
to categories 3a or 3b cannot arise simply by chance correlations longer than a single time-step and also com-
on a finite sized system. ponent sizes larger than dimers. It seems likely that some
more discriminating combinations of static and temporal
metric can further refine the behaviours observed in cate-
Nevertheless the classification clustering pattern from ex- gory 3b.
amining the 1-step time correlation function scatter-plotted
against the fraction of live cells does seem to give a good In this work we have limited the study to known named au-
combined discriminator and grouping that matches obser- tomata in the GoL family. There is scope to conduct a more
vation. systematic search using the metrics discussed here to look
for other category 3a automata. These criteria might also
The fraction of initial live cells can be varied and a variant be usefully applied to higher dimensional GoL type mod-
of figure 5 interpreted. We do not include these due to els where it is considerably harder to visualise the patterns
lack of space, but the effect is that if p is too low many of and behaviours.
the automata patterns do not have enough material to form
and so they die out with no live cells and 100% correlation
forever. References
Likewise if p is too large then overcrowding also prevents
[1] Adamatzky, A. (ed.): Game of Life Cellular Automata. No.
the complex patterns from forming. The structure shown ISBN 978-1-84996-216-2, Springer (2010)
in Figure 5 is largely stable and unchanged if the stability [2] Bays, C.: Candidates for the game of life in three dimen-
condition 0.3 < p <= 0.5 is satisfied.
 sions. Complex Systems 1, 373–400 (1987) [21] Martin, B.: Damage spreading and mu-sensitivity on cel-
[3] Bays, C.: A new candidate rule for the game of three- lular automata. Ergod. Th. and Dynam. Sys. 27, 545–565
dimensional life. Complex Sytems 6, 433–441 (1992) (2007)
[4] Bays, C.: A note on the game of life in hexagonal and pen- [22] Martinez, G., Adamatzky, A., Morita, K., Margenstern, M.:
tagonal tesellations. Complex Systems 15, 245–252 (2005) Game of Life Cellular Automata, chap. Computation with
[5] Bays, C.: Game of Life Cellular Automata, chap. The Competing patterns in Life-Like Automaton, pp. 547–572.
Game of Life in Non-Square Environments, pp. 319–329. Springer (2010)
Springer (2010) [23] Nehaniv, C.L.: Evolution in asynchronous cellular au-
[6] Binder, P.M.: (anti-) stable points and the dynamics of ex- tomata. In: Proc ICAL 2003 - Eighth Int. Conf. on Artificial
tended systems. Physics Letters A 187, 167–170 (1994) Life. pp. 65–73. MIT Press (2003)
[7] Boghosian, B.M., Rothman, D.H., W.Taylor: A Cellular [24] de Oliveira, G.M.B., Siqueira, S.R.C.: Parameter charac-
Automata Simulation of Two-Phase Flow on the CM-2 terization of two-dimensional cellular automata rule space.
Connection Machine Computer (Mar 1988), private Com- Physica D: Nonlinear Phenomena 217, 1–6 (2006)
munication [25] Owens, N., Stepney, S.: Investigations of game of life cellu-
[8] Capcarrere, M.S.: Evolution of asynchronous cellular au- lar automata rules on penrose tilings: lifetime and ash statis-
tomata. In: Parallel Problem Solving from Nature Proc. tics. Journal of Cellular Automata 5, 207–225 (2010)
PPSN VII. vol. 2439 (2002) [26] Packard, N., Wolfram, S.: Two-dimensional cellular au-
[9] Darabos, C., Giacobini, M., Tomassini, M.: Scale-free au- tomata. J. Stat. Phys. 38, 901–946 (1985)
tomata networks are not robust in a collective computational [27] Reiter, C.A.: The game of life on a hyperbolic domain.
task. In: et al., E.Y.S. (ed.) Proc. Cellular Automata: 5th In- Comput. & Graphics 21(5), 673–683 (1997)
ternational Conference on Cellular Automata for Research [28] Sarkar, P.: A brief history of cellular automata. ACM Com-
and Industry. vol. LNCS 4173. Springer (2006) puting Surveys 32, 80–107 (2000)
[10] Eppstein, D.: Game of Life Cellular Automata, chap. [29] Sipper, M., Tomassini, M., Capcarrere, M.S.: Evolving
Growth and Decay in Life-Like Cellular Automata, pp. 71– asynchronous and scalable non-uniform cellular automata.
98. Springer (2010) In: In Proceedings of International Conference on Artificial
[11] Evans, K.M.: Larger than life: threshold-range scaling of Neural Networks and Genetic Algorithms (ICANNGA97.
life?s coherent structures. Physica D 183, 45–67 (2003) pp. 67–71. Springer-Verlag (1997)
[12] Evans, M.S.: Cellular Automata - Brian’s Brain [30] Suzudo, T.: Spatial pattern formation in asynchronous cel-
(2002), http://www.msevans.com/automata/ lular automata with mass conservation. Physica A: Stat.
briansbrain.html, department of Neurology, South- Mech. and Applications 343, 185–200 (November 2004)
ern Illinois University School of Medicine [31] Wolfram, S.: Statistical Mechanics of Cellular Automata.
[13] Ganguly, N., Sikdar, B.K., Deutsch, A., Canright, G., Rev.Mod.Phys 55(3), 601–644 (1983)
Chaudhuri, P.P.: A survey on cellular automata. Tech. rep., [32] Wolfram, S.: Universality and complexity in cellular au-
Dresden University of Technology (2003) tomata. Physica D 10, 1–35 (1985)
[14] Gardner, M.: Mathematical Games: The fantastic combina- [33] Wolfram, S.: Cellular Automata as models of complexity.
tions of John Conway’s new solitaire game ”Life”. Scien- Nature 311, 419–424 (Oct 1984)
tific American 223, 120–123 (October 1970) [34] Wolfram, S.: Complex systems theory. Tech. rep., Institute
[15] Hastings, M.B., Reichhardt, C.J.O., Reichhardt, C.: Ratchet for Advanced Study, Princton, NJ 08540 (6-7 October 1984
cellular automata. Phys. Rev. Lett. 90, 247004 (2003) 1985), presented at Santa Fe Workshop on ”A response to
[16] Hawick, K.A., Scogings, C.J.: A minimal spatial cellu- the challenge of emerging synthesis in science”
lar automata for hierarchical predator-prey simulation of [35] Wolfram, S.: Twenty problems in the theory of cellular au-
food chains. In: International Conference on Scientific tomata. Physica Scripta T9, 170–183 (1985)
Computing (CSC’10). pp. 75–80. WorldComp, Las Vegas, [36] Wolfram, S.: Theory and Applications of Cellular Au-
USA (July 2010), www.massey.ac.nz/˜kahawick/ tomata. World Scientific (1986)
cstn/040 [37] Wylie, B.J.: Application of Two-Dimensional Cellular Au-
[17] Hernandez-Montoya, A.R., Coronel-Brizio, H., Rodriguez- tomaton Lattice-Gas Models to the Simulation of Hydrody-
Achach, M.: Game of Life Cellular Automata, chap. namics. Ph.D. thesis, Physics Department, Edinburgh Uni-
Macroscopic Spatial Complexity of teh Game of Life Cel- versity (1990)
lular Automaton: A sSimple Data Analysis, pp. 437–450. [38] Yey, W.C., Lin, Y.C., Chung, Y.Y.: Performance analysis of
Springer (2010) cellular automata monte carlo simulation for estimating net-
[18] Kaneko, K., Akutsu, Y.: Phase Transitions in two- work reliability. Expert Systems with Applications 36(5),
dimensional stochastic cellular automata. J.Phys.A. Letters 3537–3544 (2009)
19, 69–75 (1986)
[19] Kunkle, D.R.: Automatic Classification of One-
Dimensional Cellular Automata. Master’s thesis, Rochester
Institute of Technology (2003)
[20] Lynch, J.F.: On the threshold of chaos in random boolean
cellular automata. Random Structures & Algorithms 6(2-3),
239–260 (March-May 1995)