Complex Sy st ems 1 (1987) 169-180

Compet ition of Cellular Aut o mata R u les

Dav id B . B row n
Department of Zoology, University of Toronto,
25 Harbord Street , Toronto , On tario M5S lAl, Can ada

Ab stra ct . Biological phenome na of competition and select ion can be

applied to cellular automata. Subrules compete in t he presence of a
master rule in a hierarchy and resu lts of this competition are shown
to be largely independent of the master rule choice.

1. Introduction
A fundame ntal mechanism in the evolution of biological systems is compe-
tition. The compet itors are individua ls and the fittest survive to dom inate
the population as a whole. Each organism exh ibits a complex range of be -
h avior that is selected as a single coupled behaviora l set. Competition an d
select ion can be dr awn int o cellular automata th eory by combining severa l
automata rules int o a hierar chy and allowing them t o compe te for cellular
space .
Com pe tit ion t akes place in region s where different rules ar e locally ad-
jacent and selection is based upon a master ru le. Results are obtained by
t he methods of expe rimental mathematics.
Th.e local interaction of disti nct ru les through the mediation of a master
ru le can be expressed as a sing le rule over the same neighborhood with
a larger number of values . T his suggests that the behavior of complex
cellu lar automata ro les with many site values might largely be determined
by competition between simpler automata rules with fewer values.

2. A master rule
A hierarchy of rules is formed through the generalization of rule and site
value interact ions. It is subsequently restricted to be cons istent with pre-
vious cellular automata studies and allow an investigation of competition
and select ion. Consider a one dimensional cellular automaton th at has a
finit e number of sites connected in th e sha pe of a r ing. Ea ch site has a
limited range of discret e values. The evolu tion rule is a t ime invar ian t
ma pping whos e domain is all possible values of the site and its immediat e
neighb orhood of size r and whose ran ge is the set of all possible values k.

@ 1987 Complex Systems Publications, Inc.

170 David B. Br own

D ifferent evo lut ion rules can be permanently an chored to specific sites.
In t his situat ion, t he value of t he sites would evolve t hrough the mapping
of the rule at t ached to that site according to t he values of t he site an d
its immedi ate neighbo rs. Selection can only apply to t he site values . Al-
ternately a second rule can be introd uced in suc h a way t hat competition
for site occu pan cy is allowed between the two rules. The meta ru le gov-
erning the ru le comp eti tion for sites is called the "maste r rul e". Possib le
for ms of the master rul e are rich and varied . T wo restricti ons t hat are
t raditionally associated with some cellular auto mata can be expressed in
several forms in the hierarchical system. Here, restrictions are chosen to
a llow selection through competition and yield beh avior locally consistent
with the one dimensional totalistic rule sys tem desc ribed by Wolfram [11.
The evo lution of a site value is determined by the values of the neigh-
boring sites. By extension , the evolut ion of r ules over sites is com p let ely
governe d by the site rules and values in t he neighboring region . The "nu ll"
°
con figurat ion (gro und state) is a st ipulation that F IO .O.... O] = (see Wol-
fram [I D. The condition pro d uces a bias in site va lues and is use d in t his
investigation as a distinction between the dead an d living. Fitness is ex-
pressed as t he pot enti al s urvivorship of r ules wh ose site value is one. New
rules are chosen for all sites at each it erati on of the automaton . Selection is
a llowed to op er ate by choosing the site rule from a set of neig hbo ring rules
whose site value was one during the proceed ing time step.

3. Notation and formalism

a!t) is taken to den ote the value of sit e i in a one dimen sion al ce llu lar
a ut om at on at time ste p t . Each site va lue can be an int eger in the range
o through k- L and represents one of k possible values. F is an arbitrary
function given by t he mapping:

(3.1)

R}t) is an element of F wh ich is taken to denot e the specific rule at site i

at t ime t. The bin ar y pair xl') = (a!'), R,<')) describes the value and rule
associated with ever y site i of t he cellul ar automaton at time t. The master
rule G is an arbitrary function describ ing the ru le evolution over sites and
is given by t he mapping:

,,!'+l)
..I"'" = G(X(') • .i') )
,-r' ..., }{i+r (3.2)

G most gen erally has a ll F as its range and a ll 2r + I- sized se ts of all va lues
k and a ll F as its domain.
The site value a nd rule evo lution of the ce llular automaton is given by
t he linked binary m apping:

(3.3)
Competition of Cellular A utomata Ru les 171

where
.,(1+1) = G(X(I) X(I) ) (3.4)
..... I-r ' .. . , ,+r
Certain rest rict ions can be applied to keep the size of the rule class G within
reasonable bounds and provide control over the competit ive and selective
process.
A r ule chosen to occupy a site at t he next step of time evolut ion of the
site should be found in the local neighborhood of the site at the presen t time
interval. Hence, R!l} must be one of R!2
r , • • • , .R;(2r ' Restrictions sho uld be
imposed on rul e select ion to allow for fitness of rules based upon th eir site
values at time t. Hence, t hose sites in t he neighborhood of i whose va lue is
a at t sho uld not be cons idered to pass their rules for occupat ion of site i
at time t + 1. The class of master rules can be greatly curtailed by ma king
the decision of rule selection independent of t he specific rul e types in the
neighborhood of i.
F inally, it is important to insist that F(O, .. ,0) = is never mapped by °
G unless it is t he only ru le to appear as an input to G . This ensures two
properties. Any site with value zero will immediately assume a new non
null rule provided t hat one is locally available. No site can be forced into
inactivity when a non null rule is loca lly present. Hence G*, the rest rict ed
form of G, is defined as follows: H is an arbitrary function given by the
mapping
_ ( (t ) (t) )
m - H a i _ r, ..., ai+r (3.5)
where m is an integer from a to 2r and const rained such th at a~2r+", ;f. 0.

(1+1) _ _ .
Ri - G* - ~-r + H ( a;(t)_ r"'" (I) ) (t)
ai+ r (3.6)

and
G. ((0, R!~,), (0, R!~'+1),' .. (0, R!2,)) = O. (3.7)
The behavior of G* class master rules was examined experimentally.

4. Correspondence between competitive and noncompetitive cel-

lular automata
Assume there is a region S = {i - (r+4rs), . .. , i + (r +4rs)} where all sites
are occupied by either the null rule Fo or exactly one other rul e Fa. The
r ule F. contains the null mapping F(O, . .. , 0) = O. It follows that t he site
value evolut ion of t he subregion {i - [r + 2r.), ... , i + [r + 2rs)} over s time
steps is ident ical to t he site value evolution of a cellular automaton t hat
has r ule Fa permanently assigned to all sites; provided t hat both automata
have the same distribution of site values at time t.
This property occurs by forcing th e master rul e t o choose a new ru le for
a site from within th e neighborhood of that site. Fur t hermore , the mas ter
rul e is directed to ensu re th at t he choice is not t he null rul e should an
alte rnate rul e locally exist.
172 David B. Brown

A noncompeting cellular automaton with evolut ion rule Fa. is equivalent

to a competing automata if all the sites in the competing a utomat a have
either a null rule or exclusively, the site valu e evolution rule, F(I . Again,
Fa. must contain the null mapping. Any competing cellular au tomata with
a total of n unique k value, r neighborhood rules (excluding the null rule)
can be mapped into a noncompeting cellular aut omat a wi th n(k - 1) + 1
values and neighborhood r.

5. Q uantitative method.
Totalistic k = 2, r =2 rules we re used for the simulat ion of competing
automata. The site value evolut ion ru le at locat ion i was det ermined by
t he sum of site va lues a~~2 through a!~2 . The rules were identified by a
ru le number R N which corresponded to the decimal base representation of
the binary val ue transition table. For example, Ru le 100110 is Rul e 44 and
maps the summed site values of 1, 2, and 5 to value 1. All other summed
site values are mapped to value O.
The master rule G*m requires a rule selection choice for each of the
32 possible state patterns found in the neighborhood of site i, It can be
characterized by a 32 X 5 matrix of the form:

00000
eoooo
OcOOO
OOcOO
OOOcO

lcOOO
cOl00

10 1cO
10c 11

lc111

Here 0 represents a site with value 0, and 1 and c represent sites with value
1. c represents the location of t he rule that is chosen- to occupy the center
site i at t ime t + 1.
The master rules G* were divided into two classes . The first class
ensured t h at Rjt+l) = R!'it provided all) = 1. This scheme preserved rules
in their p laces as long as the rules kept site i in value 1. The second class of
master ru les were deemed "high migr ation ru les" and attempted to displace
ru les as mu ch as possible from t heir given locations. T he master rules often
did not pr eserve symmetry in the direction of average ru le migrat ion. The
asymmet ry was of little consequence as the tot alistic rul es are t hemse lves
symmetrical and unbiased to a direct ion of drift.
Competition of Cellular A utomata Rules 173

Pairwise competitions were carried out for most combinations of to tal-

istic rules. Each pairwise competition was usually contested twice but at
least once for each of 46 different master rules selected from a given mas-
ter rule class. The 46 master rules of a chosen class were the same for all
pai rwise competitions. Each pairwise competition started with a random
initialization of site r ules and values over a ring of 300 sites and was al-
lowed to progress throug h 300 steps of time. Resu lts of these simulations
are given in Tab les 1 and 2.

6. Quant itative r esults w it h discussion

T he outcome of a pairwi se conte st over 300 cycles can take on several
characteristics.

Both rul es can be eliminated from the ring of sites , leaving t he null
rule in place by default . This behavior is exhibited by the rule pair
(16,32 ) and is not surprising since both are class 1 cellular automata
ru les and would go to site value O.
One ru le can be eliminated and the other left in a locked or oscillatory
state. R ule pair (8,16) pred ictably demonstrate this trait as 16 is a
class 1 rule and 8 is class 2. Both rules can become fixed at sites,
locked in a fixed or oscillatory state. This wou ld be expected by rul e
pa ir (8,24) because they are both class 2 rules.
One rule is eliminated and t he other persists over all sites and remain
in a random cycling of site values. Such behavior predominates the
overall outcome of rule pair int eractions and cou ld well be expected
by the pair wise interaction of a class 3 ru le wit h any other class. Pair
(14,46) demonstrates t his behav ior (see F igure 1).
Both ru les could pe rsist changing both sites and values at random.
Such behavior wou ld be expected in the interaction of a rule with
itse lf such as (46,46) . Alternately it might be expected through the
interaction of two separate class 3 rules as (26,46). Such interactions
probably imply that insufficient time has elapsed to det ermine the
dominant rule (see Figure 2). Certain properties of competing rule
pairs are found repeatedly throughout the simulation results.
A particular ru le in a two rule pair excludes the other over vir t ually
all the master ru les exp lored from both the low and high migration
master ru le classes. Such behavior suggests a vast majority of rule
pair domina nces are independe nt of variations to the master rule.
Furt hermore, t hose cases such the p air (10,38) whic h shows ru le 38
winning 40% of the time , also indic ates that rule 10 never wins.

Ru le 46 by winning or enduring over all other r ules appears to be t he

most dominant of the k = 2, r = 2 totalist ic rul e set. Ru le 30 is almost
174 Davi d B. Brown

m.
,c ,. "
• ro
" .. " . ae aa ,. ,. ,.
,.,.
.. .. ,
ce

•• ,.
eo sa
"
,. .,
rc•
......" .. " "
.,,

"
,.
"
,. .. " " .. ,.
,.
.,
aa
..
,.
"
,.
•
,.
,. "
...
eo
. .. .. .... ,. ,.
"
.....,
,. " ., ., ,. ..
.. ", ..
" ..
,.

. ,.
.."
.."
.,
ee
"
" ., ,..
" .
.
"
" "
.. .,

m. ,, ,. ..,. .. ,. .. ., .. .. .. " .,, .. .. ..

. ,. .. .. ,." "., .,."
au
" •
,.
•• ,. es

" ..
.,.
•
, "... ." .."
e
.,
..
rc
.
.... .." " ". ,. ,.,. ., . ..,• .,,.."
ra ro ea
re
io
"
,.
ao
,."" ,.eo
..•
eo
,.
,.
.... .. " ,.""
.,"" •
,. .... .. .,
• ,.

". ..
"" ,.
..."
"., •
....., ,.
.. .,. ,. ., " "
, • ar

... ... .. ., .. .. ... , .

......
so

"
rc
.,

Table 1: Dominant rules . Rules described by the columns dominate

over rules described by the rows. An ins ta nce of dominance implies
that one rule persists with value 1 representa tion for a longer period
of time than the other rule and does so within th e first 200 iterat ions
of th e automaton. In all tables, rule pairs are the coordinates of col-
umn and row entries. The ru les are specified by bin ary rule numbers
and their base 10 code number. The right hand bit of t he binary
represent ation is dropp ed . 0% is represented by a blank and 100% is
repr esented by a *.
Competition of Cellular A utomata Rules 175

...•
,. ,. ,.
,,
.." .,
~,. ra u
" " " ", "
" "
••
• .. .. u

""
..
" " . ..
"aa"
""
es

"
"
" " "
"
" .. "

"ac "
""
"""
" ". "". "
" . "
"
..."" " "

.. . ..." .." . •
"
"
• u
.
.." " u
""
"

,, ea .. .. .. .. .. ..
.. .... .. ...
~ ,.

" " " "" " " "u "

•••
" " . "
" " "" ""
... .
" ....
""u " "
.. .
eo
re
""
" .. " "u
•
.., .. .
• . rr
"
,." ..
ae
eo
ea " u .."
"
•
"

..
ae

" .
ac
i s
." -a

".." "" "" ."..

" " "
.....
""" "" "
•
u
, ",, ", "
. ""
....
"
., . ,
se
"" ", .. "
" .."" .."• "
,,
,
•
"• "
•
" " " "
"

Table 2: Dominant Rules (high migration rate) . Rules described by

the Columns dominate over rules describ ed by t he Rows. An instance
of dominance implies th at one ru le persists wit h Value 1 represent ation
for a longer period of tim e th an t he other rul e and does so with in the
first 200 iterations of the automaton.
176 David B. B rown

Figure 1: Class 3 ru les which dominate. In figure I, the thin dot is

rule 10, and the thick dot is ru le 18. In this and all other figures, the
ho rizontal axis represents site position and the vertical axis represents
time increasing downwards. Sites with value 0 are blank; those with
value 1 are represented by a thin or thick dot depending upon t he ru le
at the site. Figures depicting the be havior of homogeneous reg ions of
rules are given by Wolfram (see references ).

Figure 2: Class 3 rules that do not immediately predominate. T he

thin dot is rule 26, and the thick dot is rule 46.
Com p etition of Cell ular A utomata Rules 177

as good. Corres po nd ing ru les 44 an d 28 also do weB, as do rules 38 and

22 following slightly behind. Notab ly rule groups 40, while 24 and 20
do ba dly. There app ears to be a st rong correlation of the dominance of
a rule to the numb er of different rule summat ions, it ca n map to value 1.
However there also appear s to he a correspondence with how the site valu e
summat ions t hat map to value a are interspersed with t hose t hat map to
value 1.

Examination of differ ences in simulation results incurred by switching

from low to high migration simulation classes su ggests that master ru les
play some ro le in t he dominance of automaton ru les. Whereas rul e 44
is quite dominant to rule 28 under low migration master rules, it is only
slight ly domin ant under high migration ru les.

A rul e disp ersed from a region where it predominantly occup ies numer-
ous sites, is likely to be eliminate d by th e swamping effect of ot her rules
of a different typ e in the surrounding neighborh ood . This effect appears to
lead to t he clustering of regions of r ules .

Rules cou ld random walk to extinction on a ring of finite size long

before a selection process has a noticeable effect . High migration should
tend to retard the agglomeration effect and be more conducive to potential
branching of spatial rule distributions. The difference between high and low
migration master ru les can be margina lly discerned t hough a hig her num be r
of rules winning in competition against t hemselves at low m igrat ion rates.

Many of th e very high nu mber ed rul es, 54 through 62 , te nd to fill space

with 1 value s and block up in class 2 typ e behavior . These rules h ave some-
what unstable state and space configurat ions t hat can be eas ily dislod ged .
The mix ing of class 3 rules with t he high numbered rules can repe atedly
destabilize t hem and t he result ing join t evolution of t he rul e pair suggests
class 4 beh avior . The master rule appears to play an important role in de-
termining both the dominant ru le in the pairing with high numbered rules
and the characteristic behavior of the interface between regions of ru les (see
Figures 3a and 3b) .
In contrast, some low and med ium numbered ru les are incapable or
limited in t he promot ion of sites to va lue 1. Cou p ling wit h ot her class 1, 2,
3 an d 4 ru les ca n provide th e addit ional circumstan ces necessary to rem ain
pers istent over the ring of sites . (4,16) represent a class 1 pair of rul es that
can jo intly stabilize themse lves. Class 4 rules are remarkable in t hat t hey
are often capable of ma intaining an d increasing persistence in the presence
of class 3 rules (see Figures 4a and 4b).

Certain low number labeled class 2 rules are able to predominate over
class 3 ru les by forming membrane type structures which can lead to elim-
inat ion through containment. The pair (4,24) is an example of such a
process (see Fig ure 5).
178 David B. Brown

Figure 3: Rules whose dominance is affected by the mast er rule. The

th in dot is rule 14, and the thick dot is rule 54. (A) and (B) represent
different master rules.

Figure 4: Rules that are not substantially affected by other rules. The
thick dot ie rule 20 in bot h (a) and (b) . The thin dot in (a) ie 50, and
the t hin dot is 12 in (b) .
Competition of Cellular A ut omat a Rules 179

~"
,• • • . ra
" " " ac
"
,. ae ,.
•• ..
..
ra
• " "
.,
"
,"". "
,".
"
..
to

"
"
e"o
,".
....
"
.,"
"
sa ..
..".
" "

,.
..""
,
,
•
e "•
"• "",
• • "•
"
" ,
" sa
"
..
"

ea ,. . ., .. .. . so .. "• ..
" .., ."
~ "e " " " "
a •
..
••
• " .." .. " "".. ..
, ., ""
.."" " .. er

,.
"
ac
aa
.
"
" to ..
•
. " ,.".,, .." "".
"
'ra
eo

eo

, .. "
""en
,.aa
..
." .. .
""
., eo

.... " ..
ro

" "., ..
"" " " ., "
., " "
.. .. ..". ...""
...." "
•
"i s • "".... •
...• ". .."" ""• .
"

Ta ble 3: Rules that become fixed or cycle. Rule pairs given by t he

row and column lab els descri bes th e percentage of rules t hat become
locked in a fixed or cycling pattern . T he upper triangular group de-
scribes the locked contests for t able 1 and t he lower triangular group
describ es th e locked contes ts for table 2.
180 David B. Brown

~.

F igure 5: Rules th at can block other rules by forming barriers . The

thick dot is rule 24 and the thin dot is ru le 34.

Acknowledgements
The aut hor wou ld like to thank R . Kap ral for his ass istance and cri t icism in
developing the framework for this analysis; R . Hansell, J. Palohe lm o, and S.
Findl ay for t he ir suggesti ons concern ing t he preparation of t he manuscript,
an d S. Wolfram for his help in revision of the same. This work was fund ed
in whole by an NSE RC op erating grant to R . H ans ell.

R eferences
[1] S. Wolfram, "Universality and Complexity in Cellular Automata", Physica
100 (1984) 1-35.