See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/314875773

Cellular Automata

Chapter · March 2017

DOI: 10.1002/9781118786352.wbieg0578

CITATIONS READS

8 1,911

1 author:

Andrew Crooks
University at Buffalo, State University of New York
224 PUBLICATIONS 8,343 CITATIONS

SEE PROFILE

All content following this page was uploaded by Andrew Crooks on 13 March 2017.

The user has requested enhancement of the downloaded file.

 surrounding neighbors) and the five cell “von
Cellular automata Neumann” neighborhood (which includes the
cell in question and its four surrounding neigh-
Andrew Crooks bors) along with a variation of one of these, the
George Mason University, USA extended 5 × 5 Moore neighborhood.
An important feature of CA is that the automa-
Over the last few decades our understanding and ton’s location does not move; the automaton
modeling of geographical systems have changed can only change its state. This is different
from looking primarily at aggregate patterns from an agent-based model where mobility is
to examining individuals, and from assuming often key (see Crooks and Heppenstall 2012 and
equilibrium to looking for far-from equilibrium Agent-based modeling). The position of the cells
through the lens of complexity science. One and their neighborhood relations remain fixed
way of exploring how aggregate patterns emerge over time. Information is explicitly exchanged,
from the bottom up (i.e., individuals) is through or “spread” only through neighborhoods (i.e., a
the use of cellular automata (CA) models. A CA local diffusion process). This notion of neighbor-
model is an arrangement of individual automata, hoods fits well with one of the tenets of geogra-
or a software representation of a self-contained phy, that of “everything is related to everything
entity, that is situated within a regular tessellated else, but near things are more related to each
space, such as the rectangular grid shown in other” (Tobler 1970). Additionally, traditional
Figure 1. An individual cell represents a discrete CA models only have one attribute; for example,
spatial confine of the automata, and an individual a cell could be occupied or unoccupied, but the
automaton is understood as neighboring some cell cannot contain multiple attributes such as
other automata. Each automaton processes inputs building type or date built. However, over time,
such as the characteristics or “state” information this assumption has been relaxed to create more
(e.g., whether a cell is alive or dead) from its expressive models (as is discussed further later).
neighbors and itself, and then decides whether to To illustrate how CA models work, Conway’s
change its own state (e.g., from dead or alive), in Game of Life is used. This is a simple model that
light of instructions programmed within itself. It operates over a rectangular lattice with square
is important to note that in traditional CA mod- cells. A cell has only one attribute and one of
els, all cell states are transformed by the same set two states, either dead or alive. The rules of the
of rules. The neighborhoods may take on a num- model are remarkably simple: each cell checks
ber of different configurations: agglomerations the states of itself and its nine surrounding cells
of adjacent cells are defined by their distance (i.e., Moore neighborhood) and, depending on
from an individual automaton. Two standard the states of these cells, the cell will do one of
neighborhood configurations are illustrated in three things. If the current cell state is alive and
Figure 1: the nine cell “Moore” neighborhood two cells in its neighborhood have a state that
(which includes the cell in question and its eight is alive, then the state of the cell remains in the

The International Encyclopedia of Geography.

Edited by Douglas Richardson, Noel Castree, Michael F. Goodchild, Audrey Kobayashi, Weidong Liu, and Richard A. Marston.
© 2017 John Wiley & Sons, Ltd. Published 2017 by John Wiley & Sons, Ltd.
DOI: 10.1002/9781118786352.wbieg0578
CELLULAR AU TOMATA

that of complex patterns emerging from simple

5 × 5 Moore
neighborhood
and well-defined lower level interactions via
3 × 3 Moore
neighborhood well-specified rules. The attractiveness of using
CA models relates to why there is a need to
model more generally. Specifically, the fact that
5 cell von
Neumann
the world is infinitely complex and trying to
neighborhood
Key
model all of its behavior is impossible. Modeling
= Automata allows researchers to unravel this complexity and
examine key elements of a system that are of most
interest, to test different ideas and theory, and
Figure 1 2-D cellular automata neighborhoods. to help understand and predict past and future
events. CA models allow us to do just this within
the environment of a computer. Once the model
alive state, otherwise it dies. However, if the cell
has been created, it can be run to see how the
is currently dead, it will become alive if the state
of three neighboring cells is alive. Lastly, the cell system evolves through the interaction of cells.
will die if it has fewer than two neighbors with a
cell state of alive or if more than three neighbors Background
are alive. Figure 2 shows these rules in action.
Specifically, Figure 2 shows how dead cells
(white) become alive (black) depending on The history of CA dates back to the begin-
the number of alive cells around a specific cell ning of computer history when Alan Turing
in question. Some patterns are stationary or demonstrated the idea that computers could
unchanging over time, such as the first four “reproduce” themselves, but it was not until
patterns on the left of each image where the the 1950s and 1960s that the field became
same cells stay alive, while others oscillate, for alive, so to speak. This earlier work relates to
example the three patterns on the right of each that of John von Neumann and the theory of
time slice. Others translate across the space, such self-reproducing automata, which described the
as the “glider” in the bottom of each image, so conceptual principles of a machine that was able
named because the collection of live cells appears to self-replicate (similar to the Game of Life
to slowly move across the grid. Such a model example above). Such work laid the foundation
demonstrates that from simple, well-defined of CA models today. Interested readers wishing
rules, more complex patterns can emerge and to know more about the foundations of CA are
persist through space and time. In essence, each referred to Benenson and Torrens (2004) and
cell processes information from external sources Cioffi-Revilla (2014).
and reacts to these sources based on rules. This While the discussion above has highlighted
notion will be revisited in the applications section how 2-D CA models work, CA models can also
of the entry. The key message is that depending be developed in other dimensions, such as one
on the transition rules and the current state of and three dimensions. Initially, many CA were
the cell, different complex patterns can emerge. developed on a 1-D space. For a 1-D model, cells
CA models have become a standard way of are arranged along a line where the right-hand
modeling of complex systems because they side is joined to the left-hand side to avoid edge
offer a way to model the idea of emergence, effects. Wolfram (2002) notes that for 1-D CA

2
 CELLULAR AU TOMATA

models, only 32 different rules can be used them attractive to geographers as self-similarity
with respect to cells changing their state. This is is seen in many geographical settings such as
because there are only a finite number of dead cities (see Batty and Longley 1994, and Fractal
and alive states for a cell and its two neighbors. analysis). Moreover, 1-D CA models do not just
For example, Figure 3 shows an implementation generate interesting patterns but have been used
of Wolfram’s “rule 30” (Wilensky, 2002), where to explore many social science questions, includ-
a specific pattern emerges from an initial seed ing the emergence of residential segregation or
cell in the middle of the image. how new political actors and empires emerge
This rule states that a cell becomes alive (white over time (Axelrod 1995). However, to apply
in Figure 3) or dies (black) depending on the cells CA to these systems it must be assumed that
to the right and left of it. More specifically, the each cell is an agent, and that all agents employ
cell state at t + 1 can be written as: the same rules. Such models also assume that
agents will not change their behavior and that
Time (t) Cell state their behavior is based on what has occurred
t=0 AAA AAD ADA ADD in the previous iteration, which some authors
t+1 D D D A question as being unrealistic of human systems
t=0 DAA DAD DDA DDD (Miller and Page 2007).
t+1 A A A D Moving to a 2-D view could be considered
more suitable to geographical questions given
where A is alive and D is dead. For example if their spatiality. Specifically, how objects and
the current cell is dead (D), and both its left and spaces relate to one another can be explored,
right neighbors are alive (A), the cell dies in the as CA allow for diffusion and interaction (Batty
next iteration. 2005). However, Wolfram (2002) notes that
Such simple rules also form fractal patterns, by adding additional dimensions there is no
in the sense that the patterns repeat them- remarkable difference in the occurrence of
selves across many scales due to their growth complex phenomena. Several things change,
rules reflecting successive applications of local however, by adding an extra dimension. First is
neighborhood principles and, as such, make the increased number of possible rule sets and

Time (t) = 0 t +1 t +2

t+3 t +4 t +5

Figure 2 Example of cells changing state from dead (white) to alive (black) over time depending on the states
of their neighboring cells.

3
CELLULAR AU TOMATA

Time (t)= 0 t+ 1 t +2

t+3 t+ 4 t +20

Figure 3 A 1-D CA model implementing “Rule 30,” where successive iterations are presented below each
other.

the ability to have different neighborhoods (e.g., application has been in computer graphics, such
instead of just right and left in 1-D CA, allows as the modeling of clouds in flight simulators.
the von Neumann and Moore neighborhoods). However, there are a number of 3-D CA models
While the vast majority of 2-D CA models developed to explore geographical problems,
operate on a square lattice, CA models have such as the infiltration of water to that of river
been developed on hexagonal and triangular deposits (Karssenberg and De Jong 2005).
lattices (Cioffi-Revilla 2014). However, such
delineation of space has not gained popularity
within the geographical community. On the Applications of CA
other hand, there are a growing number of
models that use a Voronoi tessellation of space CA models did not begin to feature much
(Benenson and Torrens 2004). Such spatial parti- in the geographical literature until the 1990s.
tions have been developed due to criticisms that
That is not to say that CA models were not
geographical phenomena, such as urban growth,
being developed before that time. Some of the
do not always fit the regular geometric partition
earliest models were developed in the 1960s and
of space of a square lattice. Unlike in the CA
1970s. For example, Chapin and Weiss (1968)
models introduced above, the form of the neigh-
borhood and the number of neighbors varies developed a cell space model to explore urban
between automata in the CA. For example, growth in Greensboro, North Carolina. While
neighborhoods using Voronoi tessellations can not adopting the term CA, their model used a
be based on shared edges as shown in Figure 4: grid to represent the settlement, and the attrac-
cell 258 has neighbors 261, 264, 266 and 274, tiveness of a cell to become urban was based on
while cell 266 only has one neighbor, that of 258. adjacency relationships. With a CA-like model
CA models have also been developed in that generated a movie of growth in the Detroit
3-D and are sometimes referred to as voxel area, Tobler (1970) showed how land-use could
automata. In such models, cells are organized be defined using a neighborhood and how
into identical cubes or voxels, and the Moore land-use could change over time. Such models,
and von Neumann neighborhoods are extended along with others, showed how one could model
in the third dimension. The predominant area of socioeconomic processes at a fine scale.

4
 CELLULAR AU TOMATA

265

264 223
261 209

274

258 214
(Neighbors: 261, 264,
266, 274)

266
554
(Neighbors: 258)
(Neighbors: 528,
545, 555) 528

191

781

544
555 536

545
193
780

0 1 2 3 4 5 Km

Figure 4 Voronoi tessellations of space where each polygon has a different number of neighbors based on a
shared edge.

The rapid growth in applications relates remotely sensed images and allows a one-to-one
to two broad advances within the sciences mapping between cells within the CA model and
more generally. Specifically, the evolution of remote sensing. An example is the DINAMICA
complexity science and the digital revolution, model (Soares-Filho, Cerqueira, and Pennachin
when computing capacity and graphics became 2002), which was designed to simulate landscape
more accessible to researchers combined with dynamics in the Amazonian tropical forests: its
the release of remotely sensed imaging (Batty cell transition rules are based on spatial data
2005). The CA model regular lattice structure is pertaining to soil types, vegetation, altitude,
analogous to the raster data structure from many slope, distance to rivers or distance to roads; the

5
CELLULAR AU TOMATA

model uses this information to explore land-use earthquakes to that of the spread of diseases, and
and land-cover change. the spread of fires. Perhaps the largest area of
Many consider the incorporation of spatial applications within the geographical community
data into CA models of geographical systems is that of land-use change. CA models have
important because this integration provided the been developed to explore deforestation due to
ability to have cells that are related to actual geo- pressure from farming to that of exploring the
graphic locations. This is of crucial importance re-growth of forests (Soares-Filho, Cerqueira,
with regard to modeling geographical systems, and Pennachin 2002) to that of urban growth.
as everything within a city, region or country is The popularity of CA in this area is due to
connected to a place. Furthermore, it allowed the fact that to truly understand and explore
modelers to think about how objects (i.e., cells) land-use change it is necessary to capture both
and their aggregations interact and change in temporal and spatial dynamics of such change,
space and time. For geographical information which is a movement away from past generations
systems (GIS) users, it provided the ability to of models, which focused on equilibrium out-
model the emergence of phenomena through comes rather than on their dynamics and which
individual interactions of features within a GIS assumed that land-use change processes were
over time and space. Moreover, some would well behaved (Benenson and Torrens 2004).
consider this linkage highly appealing in the With respect to geographical modeling, CA
sense that while GIS provides us with the ability models represent a significant improvement on
to monitor the world (i.e., land cover), it provides previous generations of simulation models: they
no mechanism to discover new decision-making are spatially implicit while contemporary models
frameworks, such as those that can be explored generally conceive of space in relative terms
through a CA model by applying rules to cell (e.g., spatial interaction and gravity models,
states. Moreover, the use of spatial data within econometric models, and location-allocation
CA models also allows researchers to validate models). CA additionally treat individual enti-
and calibrate their models. Several validation and ties or objects at a microspatial level and as
calibration techniques exist; readers are referred temporally dynamic, embodying processes that
to Pontius et al. (2008) for a detailed discussion do not imply any equilibrium end state, and
of these techniques, as it is beyond the scope of reflect simulations based on local rules which
this entry (see also Validity and verification). generate development, one cell at a time, over
The flexibility offered by CA has resulted in many time periods, all of which capture notions
this technique being used across the spectrum of from complexity science (Benenson and Torrens
geography: from that of exploring the physical 2004). With respect to defining transition rules,
world to that of exploring the human world and multiple statistical and nonstatistical methods
anywhere in between. For example, representa- have been developed. For example, Clarke and
tive applications of where CA models have been Gaydos (1998) used a Monte Carlo method,
developed to explore social science issues include while logistic regressions are favored by others
the emergence of political actors, segregation, such as Verburg et al. (2002), and nonstatistical
social interaction, and traffic dynamics. In the methods for defining transition rules include
physical arena, CA models have been developed artificial neural networks or particle swam opti-
to explore geomorphological processes, ground- mization (readers are referred to Liu and Feng
water flow and the contamination of soils, (2012) for further discussion).

6
 CELLULAR AU TOMATA

However, as with all modeling approaches, alive). Another weakness of CA models is their
CA models have their weaknesses. One is their homogenous state transition rules, which are
reliance on neighborhoods for the propagation of applied to all cells in a way that limits modeling
information, which inhibits action-at-a-distance processes that vary over space and by individual,
from taking place and makes them unable to such as human behavior. While this is not an
handle explicit spatial interaction between cells. issue for many physical systems, it is an issue for
Additionally, real-world neighborhoods come in systems where humans are involved.
many different shapes and sizes that may not map
well onto the CA tessellation. Also, when mod-
eling physical applications, care must be taken Future research directions
that the basic object being modeled is consistent
with the physical processes involved. Another In response to the various challenges faced in
weakness of CA models, specifically in relation to using CA, one area that is currently being inves-
geographical systems, is the inability of automata tigated is the linkage of CA with agent-based
cells to move within the lattice in which they models (ABM) (Heppenstall et al. 2012). In
reside. It has been argued, for example that practice, CA and ABM have often been applied
one can only go so far with representing urban separately to explore a wide variety of urban
development as fixed locations (Batty 2005). phenomena, but they are increasingly joined.
There is a need for mobile objects that can move For example, CA models are commonly applied
between cells, for example, when dealing with to represent possible land-uses changes, while
mobile objects such as pedestrians, migrating ABM are often applied to crowd dynamics
households, or relocating firms. However, this and traffic simulation. However, models are
limitation has been overcome using the cells just increasingly being developed using a combina-
as locations in which the agents can move around tion of CA and ABM techniques to overcome
and, more importantly, can communicate with some of the weaknesses identified above and
other agents. For example, the FACS (free agents the distinction between them is increasingly
in a cellular space) models by Portugali (2000) blurred. For example, cells may be granted
show how important it is to consider mobility some degree of agency in their state descriptions
within the process of segregation. and are simply reinterpreted as artificial agents
As a result of these weaknesses, CA models and/or agents are imposed on top of a CA
applied to geographical systems rarely invoke the layer, while simulated agents are interpreted as
strict limits of CA outlined above. Researchers responding to average cell conditions and are
have adapted the formalism of CA to meet able to move location. This combination of CA
their simulation needs. For example, Clark and and ABM results in the continuum between
Gaydos (1998) relaxed the cell neighborhood the two approaches narrowing and, as such,
assumption entirely, in that cells are chosen for these models are better described as agent-based
development across the entire space without models operating in cellular space (CS) models
any constraints imposed by local neighbors. rather than pure CA models (Batty 2005). By
O’Sullivan and Torrens (2000) discuss adding combining the two approaches, the entire range
several cell states to account for continuous of action and interaction effects relevant to geo-
change in land-use and demographic makeup, graphical systems can be accessed, for example,
instead of purely discrete states (e.g., dead or action-at-a-distance. Additionally, by granting

7
CELLULAR AU TOMATA

CA models some degree of behavior, decisions dynamic complex systems; Validity and
can be taken account of; for example, in a CA verification
model, it might be assumed that a land cell will
be converted from rural to urban if its surround-
ing cells are all urban. However, ABM allows the References
behavior of individual cells to be considered. For
example, if the land cell owner does not desire Axelrod, R. 1995. “A Model of the Emergence of
to sell the land cell to a developer, the land will New Political Actors.” In Artificial Societies: The
not be developed. This individual behavior is Computer Simulation of Social Life, edited by N.
often missing in CA models of urban growth. Gilbert and R. Conte, 15–32. London: UCL Press.
Batty, M. 2005. Cities and Complexity: Understanding
One approach that uses this combination is
Cities with Cellular Automata, Agent-Based Models,
that of “geosimulation” (Benenson and Tor-
and Fractals. Cambridge, MA: The MIT Press.
rens 2004), which stresses the importance of Batty, M., and P.A. Longley. 1994. Fractal Cities: A
studying space when seeking to understand our Geometry of Form and Functions. London: Academic
environment. These involve problems where Press.
nonfixed objects (e.g., people or cars) are located Benenson, I., and P.M. Torrens. 2004. Geosimula-
on fixed objects (features such as houses or tion: Automata-Based Modelling of Urban Phenomena,
roads) allowing for the simultaneous processing Chichester, UK: John Wiley & Sons.
of immobile and mobile objects, at the same Chapin, F.S., and S.F. Weiss. 1968. “A Probabilis-
time directly accounting for spatial relationships tic Model for Residential Growth.” Transportation
between various types of urban and nonur- Research, 2(4): 375–390.
ban objects, revealing the emergence of spatial Cioffi-Revilla, C. 2014. Introduction to Computational
Social Science: Principles and Applications. New York:
patterns at higher levels of urban hierarchy.
Springer.
Clarke, K.C., and L.J. Gaydos. 1998. “Loose-
Summary Coupling a Cellular Automaton Model and GIS:
Long-Term Urban Growth Predictions for San
Francisco and Baltimore.” International Journal of
This entry has introduced CA models, presented Geographic Information Science, 12(7): 699–714.
its history with respect to geographical systems, Crooks, A.T., and A. Heppenstall. 2012. “Introduc-
discussed its application domains, and high- tion to Agent-Based Modelling.” In Agent-based
lighted future directions of research. Readers Models of Geographical Systems, edited by A. Hep-
wishing to explore the topic further are referred penstall, A.T. Crooks, L.M. See, and M. Batty,
to Benenson and Torrens (2004), Batty (2005), 85–108. New York: Springer.
Heppenstall, A.J., A.T. Crooks, M. Batty, and L.M.
and Wolfram (2002), whose books give further
See, eds. 2012. Agent-based Models of Geographical
reviews of CA principles, models and their Systems. New York: Springer.
applications. Karssenberg, D., and K. De Jong. 2005. “Dynamic
Environmental Modelling in GIS: 1. Modelling in
Three Spatial Dimensions.” International Journal of
SEE ALSO: Agent-based modeling; Geographical Information Science, 19(5): 559–579.
Complexity in biogeography; Fractal analysis; Liu, Y., and Y. Feng. 2012. “A Logistic Based Cellular
Land-use/cover change and climate; Automata Model for Continuous Urban Growth
Location-allocation models; Representation: Simulation: A Case Study of the Gold Coast City,

8
 CELLULAR AU TOMATA

Australia.” In Agent-based Models of Geographical Sys- Automata Model Designed to Simulate the Land-
tems, edited by A. Heppenstall, A.T. Crooks, L.M. scape Dynamics in an Amazonian Colonization
See, and M. Batty, 643–662. New York: Springer. Frontier.” Ecological Modelling, 154(3): 217–235.
Miller, J.H., and S.E. Page. 2007. Complex Adaptive Tobler, W. 1970. “A Computer Movie Simulating
Systems. Princeton, NJ: Princeton University Press. Urban Growth in the Detroit Region.” Economic
O’Sullivan, D., and P.M. Torrens. 2000. Cellular Mod- Geography, 46(2): 234–240.
els of Urban Systems. Working Paper 22. London: Verburg, P.H., W. Soepboer, A. Veldkamp, et al. 2002.
Centre for Advanced Spatial Analysis (University “Modeling the Spatial Dynamics of Regional Land
College London). Use: The CLUE-S Model.” Environmental Manage-
Pontius, R.G., W. Boersma, J.C. Castella, et al. 2008. ment, 30(3): 391–405.
“Comparing The Input, Output, and Validation Wilensky, U. 2002. “CA 1D Rule 30.” NetLogo.
Maps for Several Models of Land Change.” The http://ccl.northwestern.edu/netlogo/models/
Annals of Regional Science, 42(1): 11–37. CA1DRule30 (accessed February 22, 2016).
Portugali, J. 2000. Self-Organization and the City, Wolfram, S. 2002. A New Kind of Science. Champaign,
Berlin: Springer. IL: Wolfram Media.
Soares-Filho, B.S., G.C. Cerqueira, and C.L. Pen-
nachin. 2002. “Dinamica – A Stochastic Cellular

View publication stats