Fractals, Vol. 9, No.

2 (2001) 177–184
c World Scientific Publishing Company

FRACTAL METHODS APPLIED TO DESCRIBE

CELLULAR AUTOMATON COMBAT MODELS

MICHAEL K. LAUREN
Defence Operational Technology Support Establishment,
Private Bag 32901, Devonport Naval Base, Auckland, New Zealand
E-mail: [email protected]

Received August 1, 2000; Accepted November 16, 2000

Abstract
Analysis of warfare data provides compelling evidence that intensity of conflicts obeys a power-
law (fractal) dependence on frequency. There is also evidence for the existence of other power-
law dependences and traits characteristic of high-dimensional chaotic systems, such as fat-tailed
probability distributions and intermittency in warfare data. In this report, it is discussed how a
cellular automaton model used to describe modern maneuver warfare produces casualty distri-
butions which exhibit these properties. This points to a possible origin of the characteristics
of the larger timescale data. More interesting, the techniques of fractal analysis offer a method
by which to characterize these behaviors, and to quantify the difference between models based
on complexity theory (such as cellular automata models), and more traditional combat models
based on the physics of military equipment.

1. INTRODUCTION relationship between infantry force ratio (i.e. the

number of attackers to defenders) and attack
Application of fractal methods to describe the sta- casualties.4,5 These relationships have an interest-
tistical distribution of the intensities of wars has ing and natural geometric interpretation in terms
been discussed in this journal by Roberts and of the fractal formalism.6
Turcotte.1 Their analysis showed that fractal di- New Zealand Defence Force’s Science Organi-
mensions exist for such data, confirming the obser- zation, the Defence Operational Technology Sup-
vation of Richardson that warfare statistics obey port Establishment, has been comparing the degree
power laws.2,3 to which conventional combat models versus the
Other types of power laws have also been identi- methods of complexity theory are able to produce
fied in combat data, such as a non-integer power-law statistics that resemble these results.7

177
178 M. K. Lauren

In particular, investigations have centered on a be downloaded from CNA’s Website (www.cna.org/

cellular automaton model called Irreducible Semi- isaac), which contains information on the workings
Autonomous Adaptive Combat (ISAAC).8 The last and philosophy of the model.
two decades have seen a growing interest in the use Each automaton’s behavior is governed by a
of cellular automaton models to describe a variety of set of parameters that determine its propensity to
problems, such as turbulence, forest fires,9–12 traf- move towards friendly or enemy units, and towards
fic jams,13–15 sand piles,16 and air traffic,17 partic- its objective. A further set of parameters act as
ularly after the work on “self-organized criticality” conditional modifiers to this process. A cluster
by Bak et al. Such models have also caught the parameter “turns off” the automata’s propensity to
attention of defence technology agencies.18 –22 move towards friends above some maximum cluster
The ISAAC model was developed as part of the size; another parameter prevents automata moving
US Marine Corp Combat Development Command’s towards their objective without a minimum number
Project Albert23 by the Center for Naval Analy- of friendly units accompanying it; and another de-
ses. Weapon systems are not explicitly described termines the minimum local numerical advantage
within the model, which relies on a high level of a group of automata require before engaging the
abstraction. Rather, the model is intended as a enemy in combat.
“distillation” of the key ingredients of combat. Yet A final set of parameters describes the basic capa-
despite this simplicity, the model is capable of bilities of automata, such as weapons range, sensor
demonstrating a high degree of complexity. range, movement rate, single-shot kill probability,
The failure of more conventional combat models, defensive factor and maximum number of simulta-
despite their high level of physical detail, to describe neous targets that can be engaged. The automata
the ability of units to react and organize themselves also possess the ability to communicate the position
to fit a particular situation is a serious shortcoming of enemy/friendly automata to friendly automata
in representing reality. This leaves the analyst with within their communications range.
a very precise picture of the outcome of one par- With these parameters, a wide set of behaviors
ticular battle fought in a few particular ways with can be induced. Even so, the version of the model
certain pieces of equipment modeled to some ap- discussed here is somewhat limited, and is not
proximation of reality. Indeed, a growing number intended to describe every aspect of a military
of military experts24–26 have shown frustration at operation. Already, several new models are under
the fire-power centric models which completely fail construction that will improve functionality. How-
to explore how a poorly equipped army might adapt ever, for the purposes of this report, ISAAC does
to prevail against a superiorly equipped foe, as in show interesting behavior, particularly for fluid sit-
Vietnam. uations resembling modern “maneuver warfare.”
The difficultly cellular automaton, and more Figure 1 illustrates such a generalized concept of
generally “multi-agent” models, face in gaining ac- operation. Here, the Blue force must maneuver its
ceptance in the military operational research com- way through a dispersed Red force to reach some
munity is that it is hard to justify why they should objective (at the bottom of the frame). Various
be preferred over the more comfortable and pre- elements of the Red force react to the Blue force
dictable physics-based models. In short, there needs as it comes within sensor range, and Blue in turn
to be a better theoretical understanding of the prop-
reacts as Red surrounds it. Table 1 contains the
erties of automaton models before they can gain
values of the ISAAC parameters used to produce
wider acceptance. The work in this paper outlines
this run.
how the methods of fractal analysis are proving a
The personality rules of the simulation impose
useful tool for characterizing these differences, and
structure on the distribution as automata are
furthermore, improving the understanding of the
attracted towards/repelled by enemy/friendly au-
interaction between automata entities.
tomata. This leads to groupings of automata into
a complicated hierarchy of clusters that may
“cooperate” in attacks on Blue automata.27 The
2. AUTOMATON COMBAT MODEL
degree of clustering on the battlefield can be mea-
Here, not much detail will be given of the ISAAC sured by obtaining a fractal dimension for the distri-
model. The reader is referred to Ilachinski8 for bution, using a standard box-counting technique28
more detail. Additionally, the model itself may (using just the Red forces). The box-counting
 Fractals Applied to Automaton Combat Models 179

In order to accurately obtain a fractal dimension

representative of the distributions that evolve from
a particular set of parameters, a large ensemble
of distributions must be used, and an average D
found. This technique reveals a short scaling range
for which D was estimated as 0.70 for the case
discussed.
An argument can be made that the function
describing the attrition rate should then depend
on D.7 In a fixed time interval, a Blue automa-
ton’s “sensors” on average map out an area (box)
proportional to the speed it is travelling. The den-
sity of Red automata — they are encountered — is
related to the number of boxes of this size that ac-
tually contain automata. But this has a non-integer
power-law dependence on the size of the box, and
hence on the time interval (i.e. if l ∝ t if l is the
width of the box, and t the time interval). In such
a case, the attrition rate function depends on:

∆nB
= f (kred , ∆t−D nbox ) (3)
∆t

where nB is the number of remaining Blue au-

tomata, kred is the probability of a Red automaton
killing a Blue automaton within its detection and
firing range with a single shot, nbox is the average
number of Red automata contained in just those
boxes that contain automata, and the size of the
box is the area an automaton typically maps out
in a period ∆t. If there exists an ensemble of runs
for which the distribution of forces that evolves pro-
duce a similar value for D, we hypothesis that for
this ensemble the distribution of casualties depends
only on k and  t−D , and has statistical moments
which scale like:
 p 
∆nB
∼ kq(D,p) ∆tr(D,p) (4)
∆t

Fig. 1 An evolutionary snapshot of an ISAAC run. where q and r are non-integer powers related to
D, the angled brackets denote an ensemble average,
and the ensemble is conditional on Blue encounter-
fractal dimension, D, is defined here as: ing Red in clusters of size nbox
log N
D = lim   (1)
d→0 1
log 3. PROPERTIES OF MODEL DATA
d
(where d is the width of the box, and N the number Figure 2 shows the probability density for the num-
of boxes required to cover all the automata on the ber of Blue casualties suffered by the 500th time
“battlefield”). D is a power-law exponent, since: step of the model run shown in Fig. 1. The dis-
tribution is based on 2000 runs. The important
N (d) = d−D . (2) point is that by this time step (and indeed, well
180 M. K. Lauren

Table 1 ISAAC parameters for run shown in Fig. 1.

Personalities: weighting towards . . . (threshold range = 5)

Alive/Injured Red Alive/Injured Blue Red Flag Blue Flag

Red 10 40 0 0
Blue 40 10 10 0

Meta Personalities
Cluster Advance Combat

Red = Blue 3 1 3

Attributes
Fire Range Sensor Range Movement Rate k (baseline)
Red = Blue 3 5 1 0.05

Initial Distribution (Battlefield Size 120 × 120)

Centre Position Size

Red 60, 60 30, 30

Blue 60, 105 15, 15
Red flag 60 1

0.4
Fractal case
Shoot out
0.35

0.3
Probability density

0.25

0.2

0.15

0.1

0.05

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Fractional number of casualties

Fig. 2 Casualty-level probability densities for the fractal and shoot-out cases (each distribution has a mean of 0.15).
 Fractals Applied to Automaton Combat Models 181
-1
10
before), the battle has run its course, so that fur-
ther casualties are unlikely (i.e. Red and Blue are
no longer interacting). Casualties do not necessar-
ily reach 100% (or some other arbitrary level), be-

(time to reach 25% casualties)-1

cause the Blue force often breaks through to its goal,
or otherwise avoids encountering concentrations of
Red. Thus the distribution represents the probabil- -2
10
ity of ending a run in a particular state. Moreover,
Blue only reaches a given level of casualties for a
certain ensemble of runs, and these runs represent
the occasions where Red was able to concentrate its Shoot out
Fractal case
firepower sufficiently to inflict this level. We may Red with comms
Red without comms
explore the nature of this ensemble by measuring -3
10
attrition rate in terms of time taken to reach that -3
10 10
-2
10
-1 0
10
level of casualties, excluding runs that do not. k

Figure 3 plots attrition rate as a function of k, Fig. 3 Attrition as a function of k for the ensemble of runs
calculated in this way, for the ensemble of runs with which reached 25%. Blue casualties for various cases.
greater than 25% Blue casualties. A further con-
straint on the ensemble was that this casualty level
had to be reached during the maneuvering stage of with the fractal case having a much longer “tail.”
the simulation. Because the Blue automata stick Figure 3 also includes the attrition rate as a
close to the goal once they reach it, they are vul- function of k for the “shoot it out” case. Note
nerable to further attack if Red units drift down that for the ensemble used for this case, every run
into this area. However, this is not the situation we reached 25% casualties (though in principle, some
are trying to model, so the arguments above do not cases never reach this level if, say, one side has suf-
apply. Hence where 25% casualties are reached as ficient luck that it eliminates its opponents before
a result of this drift down behavior, the runs were this happens). The value of q for this data was
ignored. approximately 0.9.
The values for q may be found by fitting straight In the previous section, it was supposed that the
lines on the log-log plot of attrition rate versus k. value of q depended on the fractal dimension of the
Here the value for q was approximately 0.25. Note distribution of forces. There appears to be evidence
that above a certain value for k, the power law for this in other studies conducted with this model.
breaks down, with the slope tending to zero. This For example, a similar operational scenario to that
represents the case where the kill probability is suf- described above consisted of a Blue reconnaissance
ficiently high that Blue is being killed almost imme- force represented by four Blue automata operat-
diately after encountering Red, so that increasing k ing against a Red counter-reconnaissance force of
has little further effect. 16 automata.34 The Red force was modeled with
By contrast, a distribution of casualties from a and without communications. Clearly, communi-
case where each side is simply lined up to “shoot cations has an effect on the degree to which Red
it out” is also shown in Fig. 2. Note that most is able to cluster to concentrate its firepower, and
conventional combat models are little more sophis- hence on the fractal dimension of the distribution of
ticated than this approach.29–33 A spread of out- forces. The q values for the two cases were different,
comes is expected after a given period of time, but with attrition rate falling less rapidly as the value
the level of casualties at that point represents a of k was lowered for the case where communications
“snapshot” of casualty levels on the way to some were on than when they were off.
arbitrary withdrawal level. The fixed time interval Equation (4) in the previous section implied that
used to produce the distribution shown in the plot the attrition function also possesses fractal scaling
was the mean time required to run the model to structure. Figure 4 shows the number of casualties
reach the same mean casualty level as for the frac- per unit time step for the fractal case and for the
tal distribution case shown in Fig. 1. Clearly, this shoot-it-out case. Here, in order to illustrate the
distribution of casualties was quite different from temporal structure of the attrition function, it was
the case where the forces had a fractal distribution, necessary to increase both the number of automata
182 M. K. Lauren
3
Fractal case
of each case. The difference in the power laws pro-
Casualties/unit time

duced reflects the difference in the temporal struc-

2
ture. For the fractal case, r(D, 2) was −0.24, while
1
for the “shoot out” case the scaling did not produce
the required straight line.
0
0 10 20 30 40 50 60
Time step
Mean 4. CONCLUSIONS
3
Shoot out The main result presented here is that automaton
Casualties/unit time

2 combat models can generate fractal distributions of

forces, which in turn affect the attrition rate of the
1 automata. In particular, the non-integer power-
law that describes the attrition rate of certain
0
0 10 20 30 40 50 60
ensembles seems to result from these fractal dis-
Time step tributions, and is extremely significant for combat
analysis.
Fig. 4 Time series of casualties and mean casualty rate for
fractal and shoot-out cases. It is tempting to see a relationship in the data
suggesting that q(D, 2)   r(D, 2)  1  D.
However, the values obtained for q, r and D were
some 10 to 20% out for this relationship to hold.
While this might simply reflect that the estimates
for the fractal dimensions were poor for this data
owing to the size of the datasets, or other equally
practical problems affecting their estimation, it
seems more likely that if this relationship does
indeed hold, it only does so for some idealized
ensemble. On the whole, the ISAAC model pro-
duces the fractal behavior described for only brief
periods of the run. The fractal distribution might
be interpreted as a case of self-organized criticality,
as has been claimed to exist for other automaton
models. However, the rule set tends to cause the
behavior to “collapse” into non-complex situations,
where for example, the automata simply form a
single clump at some endpoint.
In terms of combat modelling, the implications
of these results are that a Red force with a rel-
atively poor kill probability may do much better
Fig. 5 Scaling of the second-order moment of the time than suggested by a straight shoot-out approach
series data shown in Fig. 4.
for certain ensembles of runs. This is explained by
the tendency of the automata to cluster together to
concentrate firepower, rather than simply being the
involved to 60 on each side, and the kill probabili- result of a series of “lucky shots,” as would other-
ties of each side (otherwise, the casualty data would wise have to be the case.
be too sparse). For the fractal case, the casual- The behavior implies that for these models, tem-
ties occur in a burst-like, non-systematic manner. poral and spatial correlations are critical and fat-
By comparison, the shoot-it-out case produces a tailed distributions are prevalent. Although degree
gradually declining attrition rate as the number of of intermittency was not examined specifically in
automata falls. this report, there is also evidence that it plays an
Figure 5 shows how the second-order moments important role for this model.6,35 This may go some
of the data scale as a function of t. The plot is ac- way to explaining the occurrence of power laws in
tually the average of 40 plots produced from 40 runs real combat data.
 Fractals Applied to Automaton Combat Models 183

What is particularly useful about this behavior is 10. P. Bak, K. Chen and M. Creutz, “Self-Organized
that the attrition function may be suitable to be de- Criticality and the ‘Game of Life’,” Nature 342, 780
scribed by the same sort of fractal cascade models (1989).
that have been successfully used to describe tur- 11. P. Bak, How Nature Works: The Science of Self-
bulence and other kinds of geophysical signals.36,37 Organized Criticality (Springer-Verlag, New York,
It also raises the question of what the suitability 1996).
of cascade models to describe this sort of automa- 12. B. Drossel and F. Schwable, “Self-Organised Criti-
cality in a Forest-Fire Model,” Phys. Rev. Lett. 69,
ton model says about the broader use of cascade
1629 (1992).
models. Specifically, are cascade models simply a
13. K. Nagel and H. J. Herrmann, “Deterministic Mod-
useful statistical “trick” to generate scaling data, els of Traffic Jams,” Physica A199 254 (1993).
or are they a valid physical model in their own 14. K. Nagel and M. Paczuski, “Emergent Traffic Jams,”
right? Clearly, for the model examined here, no Phys. Rev. E51, 2909 (1995).
explicit “top down” cascade process occurs. Rather, 15. K. Nagel, S. Rasmussen and C. L. Barrett,
it is a “bottom up” process. However, a casual ob- “Network Traffic as a Self-Organized Critical
server with no knowledge of the model might spec- Phenomena,” in Self-Organization of Complex
ulate that a cascade process was in operation, with Structures: From Individual to Collective Dynam-
clusters of automata splitting into smaller clusters, ics, ed. F. Schweitzer (Gordon and Breach, London,
which themselves split into smaller clusters, etc. 1997), pp 579–592.
as the battle evolves. Perhaps, it is a matter of 16. S. R. Nagel, “Instability in a Sandpile,” Rev. Modern
perspective. Phys. 61, 321 (1992).
17. K. Lin, A. Sisti and L. Chow, “Study on Crowded
Two-Dimensional Airspace — Self-Organised
Criticality,” AIAA J. Aircraft 35, 301 (1998).
REFERENCES 18. J. T. Dockery and A. E. R. Woodcock, The Military
1. D. C. Roberts and D. L. Turcotte, “Fractality Landscape (Woodhead Publishing, 1993).
and Self-Organised Criticality of Wars,” Fractals 6, 19. C. W. Hunt, “Uncertainty Factor Drives New Ap-
351–357 (1998). proach to Building Simulations,” Signal, 53(2), 75
2. L. F. Richardson, “Frequency of Occurrence of Wars (1998).
and Other Fatal Quarrels,” Nature 148, 598 (1941). 20. A. E. R. Woodcock, L. Cobb and J. T. Dockery,
3. L. F. Richardson, Statistics of Deadly Quarrels “Cellular Automata: A New Method for Battlefield
(Boxwood Press, Pittsburgh, 1960). Simulation,” Signal 41 (January, 1988).
4. R. C. Thornton, Historical Analysis of Infantry De- 21. F. G. Hoffman and G. E. Horne, Maneuver Warfare
fence in Woods. Volume I: Main Text, DOAC R 9321 Science 1998 (United States Marine Corps Combat
(Defence Operational Analysis Centre, MoD, UK, Development Command, 1998).
1993). 22. A. E. R. Woodcock and D. F. Davis (eds.), Analytic
5. A. Hall, W. Wright and M. Young, Simplified Model Approaches to the Study of Future Conflicts (The
of Infantry Close Combat (SMICC) — User Guide Canadian Peacekeeping Press, Clementsport, Nova
and Technical Guide, Issue 2 (Centre for Defence Scotia, Canada, 1996).
Analysis, UK, 1997). 23. A. G. Brandstein, “Operational Synthesis: Apply-
6. M. K. Lauren, “Modelling Combat Using Fractals ing Science to Military Science,” Phalanx: The Bull.
and the Statistics of Scaling Systems,” Military Military Oper. Res. 32(4), 1 (1999).
Oper. Res. Warfare Anal. Complexity Special Issue, 24. J. E. Rhodes (Lt. Gen.), in Maneuver Warfare
Science 1998, eds. F. G. Hoffman and G. E. Horne
in production.
(US Marine Corp Combat Development Command,
7. M. K. Lauren, Characterising the Difference Be-
1998).
tween Complex Adaptive and Conventional Combat
25. R. H. Scales (Maj. Gen.), “Adaptive Enemies:
Models, DOTSE Report 169, NR 1335 (Defence
Achieving Victory by Avoiding Defeat,” Joint Forces
Operational Technology Support Establishment,
Quarterly 23 (Fall, 1999).
New Zealand, 1999).
26. J. F. Watson (Lt. Col.), “Living on the Edge,” Br.
8. A. Ilachinski, Irreducible Semi-Autonomous Adap-
Army Rev. 123, 19 (1999).
tive Combat (ISAAC): An Artificial-Life Approach
27. M. K. Lauren, Firepower Concentration in Cel-
to Land Warfare, CRM 97-61.10 (Centre for Naval
lular Automata Models — An Alternative to the
Analyses, 1997).
Lanchester Approach DOTSE Report 172, NR 1350
9. P. Bak, K. Chen and C. Tang, “A Forest-Fire Model
(Defence Operational Technology Support Establish-
and Some Thoughts on Turbulence,” Phys. Lett.
ment, New Zealand, 2000).
A147, 297 (1990).
184 M. K. Lauren

28. B. B. Mandelbrot, The Fractal Geometry of Nature 34. M. K. Lauren, Exploring the Recce-Counter Recce
(W. H. Freeman and Company, 1983). Scenario with ISAAC, DOTSE Report 171, NR 1349
29. F. W. Lanchester, “Aircraft in Warfare: The Dawn (Defence Operational Technology Support Establish-
of the Fourth Arm — No. V, the Principle of ment, New Zealand, 2000).
Concentration,” Engineering 98, 422 (1914). 35. G. E. Horne and M. K. Lauren, “Operational Synthe-
30. I. Bellany, “Modelling War,” J. Peace Res. 36(6), sis Applied to Mutual NZ/US Questions,” Proc. 34th
729 (1999). Ann. Conf. NZ Oper. Res. Soc. 205 (1999).
31. R. D. Fricker, “Attrition Models of the Ardennes 36. A. Davis, A. Marshak, W. Wiscombe, and
Campaign,” Naval Res. Logistics 45(1), 1 (1998). R. Cahalan, “Multifractal Characterisations of Non-
32. D. P. Gaver and P. A. Jacobs, “Attrition Modeling in stationary and Intermittency in Geophysical Fields:
the Presence of Decoys: An Operations-Other-Than- Observed, Retrieved, or Simulated,” J. Geophys.
War Motivation,” Naval Res. Logistics 44(5), 507 Res. 99(D4), 8055 (1994).
(1997). 37. D. Schertzer, S. Lovejoy, F. Schmitt, Y. Chigirin-
33. M. Kress and I. Talmor, “A New Look at the 3:1 skaya and D. Marsan, “Multifractal Cascade Dynam-
Rule of Combat Through Markov Stochastic Lanch- ics and Turbulent Intermittency,” Fractals 5, 427
ester Models,” J. Oper. Res. Soc. 50(7), 733 (1999). (1997).