Show Summary Details

Page of

 PRINTED FROM the OXFORD RESEARCH ENCYCLOPEDIA, CRIMINOLOGY AND CRIMINAL JUSTICE (criminology.oxfordre.com). (c) Oxford University Press USA, 2016. All Rights Reserved. Personal use only; commercial use is strictly prohibited. Please see applicable Privacy Policy and Legal Notice (for details see Privacy Policy).

date: 17 February 2018

Place-Based Simulation Modeling: Agent-Based Modeling and Virtual Environments

Summary and Keywords

Since the earliest geographical explorations of criminal phenomena, scientists have come to the realization that crime occurrences can often be best explained by analysis at local scales. For example, the works of Guerry and Quetelet—which are often credited as being the first spatial studies of crime—analyzed data that had been aggregated to regions approximately similar to US states. The next major seminal work on spatial crime patterns was from the Chicago School in the 20th century and increased the spatial resolution of analysis to the census tract (an American administrative area that is designed to contain approximately 4,000 individual inhabitants). With the availability of higher-quality spatial data, as well as improvements in the computing infrastructure (particularly with respect to spatial analysis and mapping), more recent empirical spatial criminology work can operate at even higher resolutions; the “crime at places” literature regularly highlights the importance of analyzing crime at the street segment or at even finer scales. These empirical realizations—that crime patterns vary substantially at micro places—are well grounded in the core environmental criminology theories of routine activity theory, the geometric theory of crime, and the rational choice perspective. Each theory focuses on the individual-level nature of crime, the behavior and motivations of individual people, and the importance of the immediate surroundings. For example, routine activities theory stipulates that a crime is possible when an offender and a potential victim meet at the same time and place in the absence of a capable guardian. The geometric theory of crime suggests that individuals build up an awareness of their surroundings as they undertake their routine activities, and it is where these areas overlap with crime opportunities that crimes are most likely to occur. Finally, the rational choice perspective suggests that the decision to commit a crime is partially a cost-benefit analysis of the risks and rewards. To properly understand or model these three decisions it is important to capture the motivations, awareness, rationality, immediate surroundings, etc., of the individual and include a highly disaggregate representation of space (i.e. “micro-places”). Unfortunately one of the most common methods for modeling crime, regression, is somewhat poorly suited capturing these dynamics. As with most traditional modeling approaches, regression models represent the underlying system through mathematical aggregations. The resulting models are therefore well suited to systems that behave in a linear fashion (e.g., where a change in model input leads to a predictable change in the model output) and where low-level heterogeneity is not important (i.e., we can assume that everyone in a particular group of people will behave in the same way). However, as alluded to earlier, the crime system does not necessarily meet these assumptions. To really understand the dynamics of crime patterns, and to be able to properly represent the underlying theories, it is necessary to represent the behavior of the individual system components (i.e. people) directly. For this reason, many scientists from a variety of different disciplines are turning to individual-level modeling techniques such as agent-based modeling.

Keywords: agent-based modeling, crime simulation, travel to crime, virtual environment, NetLogo, virtual laboratory

Historical Context

Crime occurrences are the result of individual behavior patterns (e.g. by offenders, victims, guardians, or others). Understanding the causes and consequences of different behaviors and their relationship with crime has taxed criminologists, as well as scientists in related disciplines, since their inception. Unfortunately, the “traditional” means of modeling social systems, where heterogeneous components are distilled down into homogeneous units, make it virtually impossible to draw any meaningful conclusions about their inner workings or micro dynamics (Batty, 2008), particularly at fine geographical scales.

Agent-based modeling (ABM) is a methodology that reverses the traditional, aggregate, modeling paradigm. It is an interdisciplinary field that has emerged from the study of chaos, complex adaptive systems, and computers/computational modeling (Cioffi-Revilla, 2014; Macal & North, 2010). Although the first agent-based model is often credited to the groundbreaking work of Thomas Schelling in the 1960s (Schelling, 1969)—who was later awarded the Nobel Memorial Prize in Economic Sciences, with Robert Aumann, for his work on conflict theory and social simulations (Cioffi-Revilla, 2014)—the field did not begin its rapid rise in popularity until the late 1980s and early 1990s.

Agent-based models (ABMs) evolved from a similar class of computational model called a cellular automata. Although both methods center on simulating the behavior of individual units (“cells” or “agents”), agent-based models encompass greater qualitative and quantitative features than their relatively simple cellular automata counterparts. These features—which include individual autonomy, goal-directed behavior, mobility, and greater environmental complexity (Cioffi-Revilla, 2014)—allow ABMs to be effective at representing complex and self-organizing systems (Heath et al., 2009) whose behavior is driven by individual actors and their interactions. This approach to modeling has been shown to be much more effective at simulating complex systems—where the overall system behavior is dependent on the interactions between individual units—than traditional aggregate approaches (Batty, 2012; Epstein & Axtell, 1996; Schelling, 1969). ABM of social systems, such as the “crime system,” is one of the most rapidly growing class of computational models (Cioffi-Revilla, 2014).

The popularity of the field began to rise exponentially as personal computing (both in terms of hardware and software) removed the historical computational limits on the capabilities of social simulations (Gilbert & Conte, 1995; Gilbert & Doran, 1994). A wealth of diverse models began to emerge; one of the most important being Sugarscape (Epstein & Axtell, 1996). Sugarscape exemplified the “generative” approach to modeling social systems (broadly termed “generative social science”), whereby social phenomena are not preimagined but rather emerge from the behavior of the individual agents. This ability of ABMs to “grow” (Epstein & Axtell, 1996) social phenomena is particularly relevant for modeling spatial crime. It allows the researcher to design a model according to their theoretical understanding of the dynamics of the system (for example by designing agents who respond in a manner that is consistent with criminology theory). They can then run the model to see whether these fundamental rules lead to model outcomes that are comparable to real world observations. In other words, are the theoretical mechanisms able to produce a model that is sufficiently similar to the observed real world outcome? By using ABM as a scientific instrument, it is possible to assess the “generative sufficiency” (Birks et al., 2012) of theory.

What Is an Agent-Based Model?

Despite the volume and diversity of ABMs that have been produced, there is surprisingly no definition that universally describes them. However, many authors are in broad agreement about the major features that ABMs should exhibit (Bonabeau, 2002; Crooks et al., 2008; Crooks & Heppenstall, 2012; Epstein, 1999; Macal & North, 2010; Wooldridge, 2009). These include:

  • Autonomy. Agents act independently, free from a central controller. They can control their own state and make independent decisions.

  • Heterogeneity. Agents are not usually identical, neither in their behavior nor their state. Groups of similar agents might be formed through interactions, but these groups are not predetermined.

  • Interactions. Agents can interact with each other, exchange information, and influence each other’s behavior.

  • Environment. The model contains a space that allows agents to interact. This is not necessarily a geographical space—for example, the agents might interact over a network.

  • Reactivity. Agents can sense their environment and respond to changes.

  • Bounded rationality. Although agents tend to act rationally to some degree, they should not have full knowledge of the world and hence their rationality is bounded. This allows them to make mistakes.

Although there is a broad agreement that ABMs will include some of the components listed above, the application area and research questions ultimately determine the overall form and function of the agents. For example, in their model of street-level drug markets, Dray et al. (2008) used a variety of different agents (drug users, dealers and wholesalers, police constables, and outreach workers) who all utilized different behaviors, but the heterogeneity within these groups was limited (e.g., all drug users behaved according to the same set of rules). Similarly in the burglary model published by Malleson et al. (2013), although agents were able to interact with their environment by committing burglary, the agents themselves did not interact. Both of these models are clearly agent-based, even though they do not fully account for every feature listed above.

The Structure of an Agent-Based Model

At the most basic level, an agent-based model consists of individual agents, and means of allowing agents to interact. Typically, this consists of some spatial environment in which agents can move around, but does not necessarily need to include geography (the “space” could be a social network for example). Individual agents are equipped with behavioral rules (discussed in more detail below) that determine how they should behave under a set of given circumstances. Importantly, these behavioral rules can be founded on widely accepted theories. The rules themselves can vary substantially in complexity. Thomas Schelling’s famous model of segregation (Schelling, 1969) used agents whose behavior was determined entirely by a simple count of the number of nearby agents, whereas more recent examples from the field of environmental criminology include advanced behavioral models that represent the main features of contemporary behavioral theories (e.g., Birks et al., 2012; Groff, 2007b; Malleson et al., 2010, 2013).

The process of executing an agent-based model, as depicted in Figure 1, is as follows:

Place-Based Simulation Modeling: Agent-Based Modeling and Virtual EnvironmentsClick to view larger

Figure 1: The Process of Initializing, Running, and Analyzing an Agent-Based Model.

  1. 1. To begin with, the model is initialized. This involves creating the individual agents and situating them within a virtual environment. For modeling spatial crime this environment will typically contain the features that are necessary to allow the agents to behave in a suitable manner (e.g. some combination of buildings, nodes/anchor points, or transport networks).

  2. 2. Time in agent-based models is usually represented as a series of discrete steps called iterations. Each time the model iterates, every agent is able to assess its surroundings, make some decisions, and perform some action. Therefore executing the model involves iterating it until some predetermined condition has been reached. These conditions could include a certain number of crimes having been committed, a certain amount of time having passed, or once the model has reached equilibrium; see the video online, which presents a visualization of a running agent-based model. This example visualizes a model published by Malleson et al. (2010, 2013) and will be discussed in greater detail below.

  3. 3. Once the model has finished running, the results can be analyzed. This is usually a much more complicated process than that of analyzing results from traditional statistical models. This is because ABMs can produce a huge amount of information about the different elements of the system. For example: the model might report the times and locations of crime events for aggregate analysis; it might report the detailed movement traces of all agents for an analysis of how the agents move around the space; or it might report statistics such as the average journey to crime curve or rates of repeat victimization.

It is important to note that a huge volume of work has emerged regarding the means of designing, building, and analyzing agent-based models. This short entry cannot touch on most of these important developments. For a generic introduction to building reliable agent-based models, the interested reader could refer to Railsback and Grimm (2012) or Gilbert and Troitzsch (2005) for a broader introduction to social simulation.

Representing Agent Behavior

One of the most important (if not the most important) aspects of an agent-based model is the behavior of the agents. Ultimately the results of the model fundamentally depend on the behaviors of the agents, so it is vital that the rules are appropriate. The rules themselves depend heavily on the research question—in some circumstances very simple agents will be appropriate, whereas under other circumstances it might be necessary to implement a higher degree of complexity in decision making. Human behavior is extremely diverse, which poses problems for researchers who need to define formal mathematical rules to emulate cognitive processes. Individuals can possess complex and diverse personality traits, wildly different past experiences, varying levels of knowledge, and complicated emotions that can give rise to behavior that can seem irrational at best and entirely random at worst. Attempting to simulate any aspect of this cognitive mess might seem entirely foolish. Fortunately, however, it is not necessary to simulate the entire spectrum of human behavior; indeed that would be probably be pointless to attempt. Instead, the research questions can be used to define precisely which aspects of human behavior are important to model, and to what level of detail. For example, the offender agents in the model implemented by Hayslett-McCall et al. (2008) only have the behaviors that are important in defining the key research area that the authors are exploring: that of the journey to residential burglary. The decision to burgle is affected by features such as motivation (i.e., how motivated an individual might be to commit a burglary); desirability (i.e., the extent to which an agent perceives a target as desirable); and the degree of guardianship. It is not necessary to overcomplicate the model with further behavioral complexity, as this would not contribute to a better understanding of the underlying question (Elffers & van Baal, 2008). Indeed, by thoroughly reviewing the relevant theoretical literature and empirical results, the behavior of the agents can be designed to explicitly represent the factors that are deemed to be important. Interestingly, if the model fails to represent the underlying system properly then it might be because the underlying theoretical explanations themselves are at fault (a point that we will return to later).

For models with relatively simple behaviors, simple rules are often sufficient to determine how an agent should behave in a given situation. For example, the following threshold-based rules, adapted from Kennedy (2012), illustrate how an agent who has a certain level of hunger might decide whether to concentrate on finding food or on doing something else:

IF hunger IS ABOVE hunger_threshold THEN search_for_food

OTHERWISE do_something_else

One of the most interesting aspects of ABM is that even simple rules such as the one above can lead to unexpected model outcomes. Often it is the interactions between the agents that are important, not purely the behaviors themselves. That said, models of spatial crime will typically include more advanced features than the simple threshold-based rules outlined above. Fortunately, a number of cognitive frameworks exist that reduce the difficulty of implementing relatively advanced behavioral models. Probably the most commonly used architecture is “Beliefs-Desires-Intentions” (BDI) (Bratman et al., 1988) where beliefs represent the agent’s internal knowledge of the world (i.e., its memory); desires represent all the goals that the agent is trying to achieve; and intentions represent the most important goals that the agent chooses to achieve first. Although the BDI architecture has been widely used, it has also suffered some criticism due mainly to its reliance on practical reasoning. No action is performed without some form of deliberation (Balzer, 2000) but people rarely meet the requirements of such rational models (Axelrod, 1997).

A less widely used architecture is “Physical conditions, Emotional states, Cognitive capabilities, and Social status” (PECS) (Schmidt, 2000). Schmidt proposes that it is possible to model the entire range of human behavior by accounting for those four factors. PECS has been put forward as an improvement over BDI because it does not assume rational decision making and is not restricted to the relatively restrictive factors of beliefs, desires, and intentions (Schmidt, 2000). Instead, PECS individuals are imbued with a number of competing motives (such as “clean the house,” “eat food,” “raise children,” “sleep,” etc.). Each motive has an associated strength that changes depending on the circumstances of the agent, and whereby the strongest motive ultimately drives the agent’s behavior. The factors that can influence a motive strength include the agent’s internal state (e.g., an agent with a low energy level might feel hungry) as well as other external factors (e.g., an agent who smells cooking food might become hungry even if they do not have low energy levels, or might choose to eat simply because it is a time of day at which they would normally eat). Personal preferences can also come into play, where some people feel a need more strongly than others even though their internal state variable levels are the same. Using PECS, extremely complex motives that require considerable deliberation can be included as well as more instinctive reactions that require only minimal rational decision making (such as pulling a hand away from a burning object).

In criminological studies that use agent-based modeling, the approach to defining agent behavior is diverse. Although there has been some limited use of PECS (see Malleson et al., 2010, 2013), in most studies (e.g. Birks et al., 2012, 2013; Dray et al., 2008; Groff, 2007a,b; Hayslett-McCall et al., 2008) no explicit behavioral framework is used. Instead bespoke rules are created as needed, in a similar manner to those of the threshold-based approach discussed above. These include rules that trade off the possible gains of committing a crime with the risks, procedures that allow agents to construct a memory of the places that they have visited, and algorithms that determine how they should navigate the environment. For example, the probability of an agent offending at a given location and time in the model published by Birks et al. (2012) is the product of their motivation, the attractiveness of the potential target, and the offender’s awareness of the target:

p (offend) = p (motivated)× p (attractive) × p (aware)

Representing Space in Agent-Based Models

In the study of place-based crime, the design of the virtual environment is, obviously, extremely important. The way that the environment is represented in a model and the features that are included will determine the types of patterns that the model is able to create. Geographical Information Systems (GIS) are an invaluable tool for the study of spatial crime patterns (Chainey & Ratcliffe, 2005), and the incorporation of GIS with ABMs is an important step toward allowing models to more accurately represent the spatial context that the crime system is embedded in. That said, an accurate spatial environment is not a prerequisite for a place-based crime model, and some authors go as far as to say that accurate spatial representations detract from the core aims of an agent-based crime model because the additional complication that they bring makes it more difficult to understand the underlying behavioral rules and interactions that drive crime patterns (Elffers & van Baal, 2008). As noted by Axelrod (1997), if the goal of a simulation is to better understand the underlying dynamics, then it is the fundamental model assumptions that are important, not the accuracy of the surrounding environment. Therefore the following discussion will provide an introduction to the two main mechanisms by which space is included in agent-based crime models. It will draw on two specific examples, one using an abstract environment (Birks et al., 2012) and another using a more realistic, GIS-inspired environment (Malleson et al., 2013).

Abstract Representations of Space

One of the advantages of ABM is that it is extremely well suited to exploring how the interactions and behaviors of individual people can lead to the emergence of interesting phenomena, such as crime occurrences. In many circumstances, a highly-detailed representation of the world is not a necessary part of the model. It might be sufficient to provide a relatively simple, abstract space within which agents are able to move around and interact with each other.

Arguably the most famous example of an agent-based model with an abstract environment is Thomas Schelling’s model of residential segregation (Schelling, 1969). The model concisely demonstrates how individual choices made by agents lead to the emergence of unexpected macro-level patterns. In Schelling’s model there are two types of agents, here represented by blue and red counters, as illustrated in Figure 2. Each agent is autonomous and contains a desire to live in a neighborhood (defined by its 8 surrounding cells), that has a percentage of neighbors who are of the same type. In the example illustrated by Figure 2, this preference is set to 50% (i.e. to be “content,” an agent must live in a place where at least half of the surrounding agents are of the same type as itself). Time in the model is represented by the variable t. When the model is initialized (t=0), the agents are randomly distributed across the space. At each iteration of the model, each agent examines its neighbors and determines whether or not there are a sufficient number who are of the same type. If this number is less than 50% then its preferences are not met and the agent moves to any of the unoccupied cells. As the model iterates (t=1, t=2, and so on) the agents move around until they reach a point that they are all satisfied and none choose to move (t=n). The most interesting outcome of the model is that segregation emerges even when agents have relatively low preferences for similar neighbors. In the example illustrated here, agents are content for nearly half of their neighbors to be of a different type, and yet the model results in extremely segregated neighborhoods. Therefore, although the environment is a long way from realistically representing the real world, by observing how the behaviors and interactions of agents across space play out, it is possible to draw hypotheses about real patterns of segregation; namely that segregation might occur even when people have low preferences to live next to people of a similar type. There is evidence that such mild tastes and preferences for certain neighborhood types can lead to the emergence of real-world segregation (Benenson et al., 2002), so even this simple model might be reliably representing particular facets of human behavior.

Place-Based Simulation Modeling: Agent-Based Modeling and Virtual EnvironmentsClick to view larger

Figure 2: An Example of the Results of Schelling’s Segregation Model With Similar Preference Set to 50%. Model Generated Using NetLogo Software.

The use of abstract environments, such as that illustrated above, is common in agent-based crime models—for example Dray et al. (2008) explore drug market dynamics in Melbourne and Liu et al. (2005) design a model of street robbery in the city of Cincinnati—but this section will discuss one model in particular: that of Birks et al. (2012). This model is chosen because it exemplifies the exploratory approach to ABM by directly testing the “generative sufficiency” of various environmental criminology theories in an abstract agent-based model. In other words, it answers the question: are the underlying theories sufficient to produce a system that echoes the dynamics of the real crime system? If the answer is ‘yes,’ then there can be greater confidence in the applicability/reliability of the theories; if ‘no,’ then there is evidence that they might need refinement.

The Birks et al. (2012) model attempts to simulate residential burglary and consists of a virtual environment that contains navigational nodes (a proxy for a transport network) and potential targets (houses). Every time the model is executed, the environment is randomly regenerated. This minimizes the environment-specific impacts on the results (Birks et al., 2012). The model contains one type of agent: potential offenders. The behavior of the agents has been designed so that they are consistent with three fundamental environmental criminology theories. These rules dictate how they move and encounter targets (routine activities), choose to offend (rational choice), and learn about their environment (crime pattern theory). Importantly, each of these behaviors can be deactivated, so that the agents no longer behave in line with the specification for a particular theory. If the results of the model no longer share the characteristics of real-world burglary patterns, then the model has provided some evidence that the theories are a necessary part of the offender behavior and thus reliable. On the other hand, if a theory is deactivated and the results are still plausible, then perhaps the theory is not an important aspect of the underlying behavior of the burglars.

To quantify the reliability of the model results (i.e., to determine how similar the results are to real burglary patterns) the authors compare three characteristics of real and simulated burglary data:

  1. 1. Degree of spatial clustering using the Nearest Neighbor Index;

  2. 2. Distribution of repeat victimization using the Gini coefficient;

  3. 3. Journey to crime distances using the journey to crime curve.

A model that accurately captures the dynamics of the real burglary system will produce data that share similar characteristics (as defined by the three statistics above) with real world burglary data. Ultimately the authors find that although each theory improves the accuracy of the model, rational choice has a lower influence than routine activities or crime pattern theory. Therefore although the model environment does not represent a real place, and the results could not directly be used to inform crime prevention strategies, the use of an abstract environment reduces the additional complexity afforded by a complicated (realistic) spatial environment and allows the authors to concentrate entirely on the theoretical mechanisms.

Accurate Spatial Environments

The Schelling (1969) and Birks et al. (2012) models demonstrates how ABM can be a valuable tool for exploring how behavioral theories might play out in the real world. Their spatial representations are a long way removed from the real world, but are able to represent the characteristics that are necessary to situate the behavioral theories. Although the results from models such as these add to our broad understanding about the underlying mechanisms that drive a system, they have limited direct practical value. Of course a better understanding of broad offending behavior can inform crime reduction policy, but such models cannot assist planners with understanding how crime patterns in specific streets or neighborhoods might vary with changes in policy or changes to the physical or social landscape in an area. To do this, agent-based models require a spatial environment that more realistically represents the real world. Models that fall under this category are often termed predictive; they are used to make predictions about the current or future state of a real system.

As with explanatory crime models, predictive models are also popular. For example: Hayslett-McCall et al. (2008) have an developed agent-based burglary simulation situated in the city of Dallas, Texas; Elizabeth Groff developed a number of street robbery simulations that are situated in Seattle (Groff, 2007b, 2007a, 2008; Groff & Mazerolle, 2008a); and Park and Buckley (2015) have developed a residential burglary simulation in a three-dimensional geographical environment. While any of these could be used as examples in the following discussion, the focus will be on a burglary model produced by the lead author and colleagues (Malleson, 2012; Malleson et al., 2010; Malleson et al., 2013). This model is described in some detail in order to illustrate the ways that high-resolution spatial data coupled with advanced behavioral frameworks can be used to represent the crime system in a high degree of detail and to potentially influence policy at a local level.

The aim of the burglary simulation was to create a model that was able to reliably represent part of a real UK city, configure the model so that it reflected real spatiotemporal residential burglary patterns, and then alter the virtual environment to forecast the impacts of real local authority plans on the spatial distribution of burglary. For these reasons the virtual environment needed to include all of the important features of the “environmental backcloth” (Brantingham & Brantingham, 1993). In this case, the environment included the following:

  • A transport network of roads to be walked or driven along. Transport networks are important in a geographic crime model because they restrict the agents’ movements to certain paths and influence how they navigate the city and ultimately where they can go. The Ordnance Survey Integrated Transport Network (ITN) MasterMap layer was used. The ITN consists of line objects that represent all the different types of roads, including alleyways, motorways, and pedestrianized areas. Using these data it was also possible to vary the speed that agents travel around the environment based on the transportation available to them; agents with cars were able to travel quickly along major roads but more slowly along minor, residential streets. Such travel behavior is an important way for the agents as they build up a cognitive awareness of their surroundings (discussed in the next section).

  • Individual buildings that act as the targets for burglary, the residences of the offender agents, and the locations of “social places” (places where agents might go to socialize) and “drug dealers” (locations that can be used to purchase drugs). Here, Ordnance Survey MasterMap data were used to represent the physical environment to a high degree of detail. Figure 3 illustrates some of the physical objects that are available. These data were used to create buildings with a realistic spatial footprint and to estimate factors such as accessibility (a measure if how easy it is to gain entry to the house) and visibility (the level of visibility of the house to neighbors and passersby).

  • Communities to represent the social factors that might influence a burglary decision. One of the influences of the community type will be to determine whether a property is likely to be occupied at a given time, and who might be present on the street (i.e., a measure of guardianship). Different groups of people exhibit very different routine activities, such that, for example, houses in an area predominantly inhabited by university students will have very different occupancy patterns to those that are predominantly inhabited by families with small children. In the model, data from sources such as the UK Census and other administrative data (such as the Index of Multiple Deprivation) were used to estimate these factors. Another important aspect that can be accounted for in the communities layer is that of community cohesion. There is a body of evidence that suggests that residents in cohesive communities are more likely reduce the attractiveness of properties in the area (Sampson et al., 1997) and therefore variables were created to represent concentrated disadvantage, residential stability, and ethnic heterogeneity.

Place-Based Simulation Modeling: Agent-Based Modeling and Virtual EnvironmentsClick to view larger

Figure 3: An Example of the Ordnance Survey MasterMap Topography Layer. Taken From Malleson (2010).

As a demonstration of the effectiveness of agent-based crime models with realistic transport networks, Figure 4 illustrates the difference that a public transport route can have on simulated burglary patterns. A single agent exists who can be allowed to use a simple public transport network or not. When the agent has access to public transport they use the service to travel between their home, their drug dealer, and a place that they visit to socialize. Due to the differing awareness spaces created by different public transport methods, the locations of burglaries are much more heavily clustered when the agent uses public transport. This is because they are less aware of the overall simulation area because they have not explored it as they would do if there were no public transport availability. This seemingly simple result is significant because it demonstrates how an agent-based model can cleanly account for criminology theory such as crime pattern theory and routine activities theory.

Place-Based Simulation Modeling: Agent-Based Modeling and Virtual EnvironmentsClick to view larger

Figure 4: An Example of the Impact of Public Transport on the Hypothetical Burglary Patterns Caused by a Single Burglar Agent. Red and yellow areas highlight the places with the most burglaries. When the agent has access to public transport, their routine activities are more limited to the streets surrounding transport stops and therefore—because they first attempt to target houses that they are aware of—their burglaries are clustered around the points at which they enter or leave the transport system. When they do not use public transport, they tend to build up a greater awareness of the environment and their burglary locations are more spread out.

As with other agent-based crime models such as that of Birks et al. (2012), offender agents in the Malleson et al. (2013) model have been imbued with behavioral rules that replicate environmental criminology theories. It is important, however, to include an appropriate level of complexity—too simple and the agents might be unable to represent behavior to an appropriate degree of realism, too complex and the model becomes no easier to understand than they real system that it is based on. Theory suggests that it is important for burglar agents perform realistic daily activities in order to build up an accurate picture of the virtual environment (an “awareness space”) from which they subsequently choose burglary targets (Brantingham & Brantingham, 1981, 1993; Cohen & Felson, 1979). Therefore behaviors are required that support these typical routines. In the model outlined here, the PECS behavioral framework was used to control the agents. Each agent had three different motivations, and the strength of these changed depending on the circumstances of the agent. The three motives for the agents were: drug use (a common motivation for burglary in the study area at the time); sleep (everyone needs to sleep, and this motive has the effect of pulling agents back to their homes thus creating more realistic awareness spaces); and socializing (agents have a need to socialize occasionally). Clearly this is a vast simplification of the rich variety of needs that affect a real person but they are sufficient to create realistic daily behavior and to represent the theoretical requirements.

Another important agent component is the cognitive map. As an agent navigates the virtual environment, they keep a record of the houses (and the communities that the houses are situated in) as they pass them. This allows two important characteristics of the burglary system to be included. First, the agents’ cognitive maps will be more detailed around their homes and the places they visit on a regular basis (as illustrated in Figure 4). Therefore these areas are more at risk of burglary simply because an agent is aware of the opportunities that they present. Second, the “near repeat hypothesis” (Johnson et al., 2007) suggests that following a burglary, the victim and their neighbors have a heightened burglary risk for a short time. As agents in the model remember where they have been, it is possible to increase the likelihood that they revisit recently victimized properties.

To run the model, all agents start at home. Over time, their motivations increase in size to the point that they will need to take some action to satisfy them. If the agents want to purchase drugs or to socialize, they will first need to commit a burglary in order to generate wealth. To do this, the agents consider all of the neighborhoods that they are familiar with in their cognitive map, and visit the one that they think will be the most likely to offer a burglary target. Once they arrive in the neighborhood, the search the area by examining the individual houses looking for one that is sufficiently attractive. If they do not find one after searching the neighborhood for a predetermined time, then the agent visits another neighborhood and repeats the search. This process is repeated until a target is found and they can then continue to satisfy the original goal that has made burglary necessary. The burglary process attempts to reflect literature on, for example, “optimal foragers” (Johnson, 2004) and known search patterns (Rengert, 1996).

The model was applied to examine the consequences of an urban regeneration scheme. Interestingly it found that, after regeneration, a relatively small number of houses near to the site of the regeneration were at a greater risk than their neighbors due to their location on a main road that was regularly used by the burglar agents to navigate the area. Without modeling the heterogeneous routine activities and awareness spaces of the agents it would not have been possible to begin to forecast potential policy impacts at such a local scale. A second interesting result was that the model significantly underpredicted the number of burglaries that occurred in one particular neighborhood. On discussing this result with policing partners, it became clear that in that neighborhood in particular, burglary was often motivated by social factors (such as bullying or intimidation of the victims) rather than monetary factors. Therefore the fundamental model assumptions (that burglary was primarily financially motivated) did not hold to the same degree across the whole city. This model “failure” highlighted the places were the underlying theories did not play out to the same degree in the real world. Overall, although the additional geographical complexity afforded to models such as that outlined here makes them less amenable to testing theories, the realistic representation of the surrounding geography makes them potentially valuable as a tool for forecasting the impacts of new policies on a local area.

Tools for Implementing Agent-Based Models

Place-Based Simulation Modeling: Agent-Based Modeling and Virtual EnvironmentsClick to view larger

Figure 5: The NetLogo Interface Tab (a) and the Code tab (b).

One of the greatest hurdles that has traditionally prevented the widespread use of ABM is technical. Historically, models have needed to be programmed from scratch. Not only does this prohibit their development from people who cannot write computer code, it also requires some considerable amount of time even for those who can code. Fortunately, in recent years several agent-based modeling toolkits have been developed. Some of these reduce the burden on individuals who write their own models using computer code, and others offer graphical interfaces that allow models to be created without having to write any computer code at all. While it is impossible to discuss all relevant libraries here (there are hundreds of them), three commonly used open source libraries will be briefly outlined with a view to educating the reader about the most relevant for their application. For a fuller introduction, the interested reader can refer to the documentation for the particular toolkit (all are well documented) or to the recent review by Kravari and Bassiliades (2015).

Probably the most well known ABM toolkit, and one that is regularly used in agent-based crime modeling (e.g., Birks et al., 2008, 2012, 2013) is NetLogo. The two main elements to the NetLogo program, as depicted in Figure 5, are the “Interface” tab and the “Code” tab. The Interface supports graphical elements that can be used to control the simulation or to report results (e.g., sliders, buttons, switches, and graphs). The “Code” tab is used to define the model logic (i.e., how the model should run and how the agents should behave). Although NetLogo does require the use of some computer coding, its language is based on the Logo language which was originally developed to help children to learn to code. This means it is a reasonably easy language to learn when compared to other, more advanced computer languages.

Another very popular toolkit is Repast (the Recursive Porous Agent Simulation Toolkit). Like NetLogo, Repast provides useful model components, such as functions to create graphs, take videos, and schedule events. It also has extensive functions to read and write spatial data and to create geographically-realistic worlds. Repast was used to create the Malleson et al. (2013) model discussed earlier. Models in Repast can either be created using the Java programming language, a language that is similar to Logo called “ReLogo,” or though point-and-click state charts (no programming). North et al. (2013) provide a good review and summary of Repast’s current capabilities.

Finally, MASON (the Multi Agent Simulation Of Neighborhoods) is a less well known library that uses the Java programming language. As with other toolkits, its core functionality includes dynamic charting (e.g., with histograms and line graphs), model visualization, and data input/output. The GeoMASON extension allows developers to include geographically realistic space in their model. An example of such a model is illustrated in Figure 6. Although the requirement to first learn Java is a hurdle for some model developers, the toolkit is efficient and contains extensive analysis and modeling features.

Place-Based Simulation Modeling: Agent-Based Modeling and Virtual EnvironmentsClick to view larger

Figure 6: A Spatial Model Under Development Using the MASON Toolkit.

A Critique of ABM for Modeling Spatial Crime

Thus far, agent-based modeling has been presented as a powerful methodological tool for exploring spatial crime patterns and for better understanding crime theories. There are, of course, difficulties that have yet to be overcome by the community of modelers. The following will summarize some of the benefits of the use of ABM for modeling spatial crime before discussing the major difficulties.

Appeal of ABM

The central difference between ABM and more traditional, aggregate methods, is that ABM simulates the behavior of the individual components whose activities ultimately define the behavior of the overall system. In effect, this approach to modeling systems allows models to capture certain system behaviors that are very difficult to represent with aggregate methods that disregard the importance of individual heterogeneity and interactions. A prime example here is emergence:

A phenomenon is emergent when it can only be described and characterized using terms and measurements that are inappropriate or impossible to apply to the component units.

(Gilbert, 2004)

In this sense, the “crime system” is emergent. No one individual plans how crimes should be committed across a time and place; the overall distribution of crime is an emergent outcome from the behaviors and interactions of countless individual entities (victims, offenders, guardians, etc.). Crime patterns emerge from these lower-level interactions. This “bottom-up” view of modeling systems is also an excellent means of capturing the core theoretical concepts that underpin environmental criminology. Crime theories predominantly focus on the behavior of individual people, not aggregate, homogeneous groups, so ABM presents a clean and clear way of exploring the ways that theoretical assumptions play out.

Another advantage to ABM, which is closely related to its ability to model emergent phenomena, is its ability to allow modelers to create a natural description of the system under study (Bonabeau, 2002). There are many systems, particularly in the social sciences, that cannot be sensibly modeled using mathematical equations (Moss & Edmonds, 2005). In an agent-based model, the rules that drive the behavior of the system are specified directly for each individual (exactly as they are prescribed by theory) so it is not necessary to try to coax an aggregate model into performing as if it were modeling individuals directly. In ABM the “natural variety” of cities becomes part of the model, rather than being smoothed out through aggregation (Batty, 2005; Brantingham & Brantingham, 2004). Also, the ability to implement theories directly helps to create a bridge between verbal theories and mathematical models. Natural language can be ambiguous whereas computer code is precise. Therefore in order to implement a theory in an agent-based model, the theory must be sufficiently rigorously specified in order to convert its natural language description into precise computer code. In effect, ABM can promote the “rigorous specification of theory” (Birks et al., 2012).

Difficulties

Of course there are some disadvantages to the use of ABM for modeling social systems, and spatial crime in particular. An advantage with ABM is that it is possible to account directly for “human” traits such as complex psychology and seemingly irrational behavior (Bonabeau, 2002). However, formally defining these “soft” characteristics in a computer model is actually extremely difficult. As these factors must be defined explicitly in an agent-based model, there can be a tendency toward minimal behavioral complexity (O’Sullivan & Haklay, 2000). The task of adequately defining behavior is further aggravated by difficulties in gathering appropriate data to understand complex, dynamic societies (Gilbert, 2004) (a point that will be revisited shortly).

Theories will rarely stipulate precisely how someone will behave under given conditions. Instead, they specify the likely behaviors that might take place. This probabilism can be accounted for in ABMs by applying probabilities to particular behaviors. For example, the decision to burgle a house might be specified as a probability that is calculated from factors such as the affluence of the house, taking into account the presence of security measures, the presence of guardians, and an agent’s own internal needs or drivers or both. If the probability is low, the agent will be less likely to commit the burglary, but it is still a possible outcome. Therefore in one model run the agent might choose not to commit the burglary, but in another they might take the opposite decision. This makes the outcomes from a single model run unreliable—they could be the result of pure chance—so models must be executed a number of times to create robust results (Axelrod, 1997). Birks et al. (2012), for example, ran each configuration of their model 500 times. While this is not a necessarily a problem in its own right, it means that models can become extremely computationally expensive. ABMs are usually computationally expensive anyway, especially those that involve difficult spatial operations (such as finding the number of buildings surrounding an agent), and having to run a single configuration hundreds of times can make it impossible to obtain robust results.

The ability to create a natural description of a system that ABM offers can have a potential drawback: it becomes very easy to overcomplicate a model. As Birks et al. (2012, p. 245) put it:

While ABM offers considerable flexibility, the limitless directions a model can progress in dictate that researchers must exercise considerable self-control in selecting what should and should not be included in any model.

A model that is overly complicated might be no easier to understand than the system that it is trying to simulate in the first place. Therefore the model designer needs to identify the most important aspects of the system that should be included in a model and exclude all else. Identifying these “important” aspects is not necessarily a simple task, especially if the theoretical explanations that underpin the system behaviors are poorly specified or not well understood.

Where theories are poorly specified, it is often possible to use empirical results or secondary data to parameterize a model. However, because ABM operates at the level of the individual it can often be very difficult to obtain reliable, accurate, high-resolution data on individual behavior. It is possible to simulate extremely complex behaviors using ABM, but if there are no individual-level data with which to validate the model then it is difficult to have confidence in the results. In effect, the advantages of high model complexity and flexibility are tempered by the difficulty of finding suitable values for model parameters based on empirical evidence. For example, in the Malleson et al. (2013) model, agents were able to socialize and visit drug dealers. However, there is very limited empirical evidence for where individual agents might go to undertake these activities and the best that can be done instead is to estimate based on theory and whatever aggregate empirical evidence is available. The proliferation of new “big data” sources might offer some answers here. For example, individuals are publishing data about themselves through sources such as social media on an unprecedented scale. There is some potential to utilize these sources to better understand individual behavior and to calibrate models. For a recent discussion, see Heppenstall et al. (2016). Obviously this approach has serious ethical implications which are discussed briefly below.

Finally, there are also difficulties that relate to the implementation of an agent-based model. This relates to the fact that relatively small pieces of computer code can form integral parts of hundreds, thousands, or even millions of agents. It is therefore possible that small errors in the logic of the code can have huge effects on the outcome of the model. This is further compounded when attempting to ensure that work is repeatable because when programmers implement models, they unavoidably include many assumptions which are not documented (O’Sullivan & Haklay, 2000)—although more recently, efforts are being made to overcome this last point through the development of protocols such as ODD (Grimm et al., 2006, 2010) and with model repositories such as OpenABM.

Ethical Implications

There are undoubtedly complicated ethical implications associated with the use of ABM to simulate criminal phenomena. The most obvious implication arises from the question: “If an agent-based model is good enough, can it tell me where someone will commit their next crime?” In other words, if an agent-based model is sufficiently realistic, and the agents represent real people, could it be possible to predict if and where and when they will commit crime? There has already been some (albeit limited) interest in this possibility in the media (e.g., Arthur, 2010). While agent-based crime models are a long way from such a level of accuracy, it is important for modelers to clearly set the limits to what they deem as ethically acceptable. Interestingly, there are no established ethical frameworks that modelers can draw on explicitly for agent-based modeling work, although there are attempts to begin creating some (Evans, 2012). Although standard data protection legislation will in some countries (such as the U.K.) limit academic research that utilizes individual-level data and has policy implications, there are often clauses that permit research with police data. That said, the forthcoming incorporation of the EU General Data Protection Regulation into the legal frameworks of European countries will have a significant impact on work that utilizes personal data. Either way, it is down to the modelers to make it clear that some areas of work, even if they were possible and legal, would be highly ethically dubious. Broadly, university research ethics committees are probably well suited to evaluating these risks.

These ethical risks are also becoming particularly pertinent in the “age of big data.” Sources that are typically used to derive information about individuals (such as mobile phone use, social media activity, loyalty card use, etc.) are possibly predicated on relatively weak consent. Although many people will agree to the terms and conditions when they use a service, it is not clear that they have read the documentation, let alone understood it. Similarly, if the biases in the data are not properly understood and accounted for, models run the risk of being wrong at best, and at worst exacerbating existing biases. For example, there is evidence that PredPol, the well known predictive policing algorithm used by a number of forces across the United States, inadvertently exacerbates existing biases in police data (Lum & Isaac, 2016). Fortunately some of the methodologies that underpin PredPol are publicly available, but many others are closed commercial secrets. It is imperative that agent-based crime models are open source so that the assumptions can be clearly understood and there can be no confusion about what the results show, the context in which they were created, and the limits to their applicability.

Conclusion

Crime occurrences are the result of the interactions between individual people with their own unique motivations, behaviors, cognition, etc., and are located in a geographical environment that can act to inhibit or even catalyze crimes. Agent-based modeling is an approach to modeling systems that focusses on the individual components that ultimately drive the overall behavior of the system. In an agent-based model, virtual “agents” are created who typically represent people. In the context of crime modeling, these are often “offenders,” “victims,” “guardians,” etc. These agents are placed in a virtual environment that can reflect the real world to a greater or lesser extent as necessitated by the research question. Such environments range from the abstract (i.e. a flat, regular, grid-like space) to geographically realistic (i.e. with roads, buildings, etc.). The choice of environment is dependent on the underlying research question. The agents themselves have behavioral rules that determine the actions that they should take in a given situation, and as a model executes the agents are able to observe their surroundings, make decisions, and perform actions. In this manner, a model can re-create the system under study by modeling the actions of the individuals from the bottom up, rather than trying to construct aggregate equations to represent the system from the top down.

While not widespread, the application of ABM is growing and shows great promise as a means of both exploring criminology theory and for answering empirical questions, particularly when geography and the surrounding environmental context are important. There are complicated ethical implications that must be considered, but if conducted carefully there is no reason that such models cannot continue to be a useful tool for criminologists and practitioners.

The NetLogo website lists a large number of agent-based models that can be run either by downloading the software, or directly in the web browser:

NetLogo.

Repast and Mason can be found the following URLs respectively:

Repast.

Mason.

Further Reading

This entry has necessarily ignored a large number of important methodological aspects associated with agent-based modeling. For the reader who would like to learn more about ABM in general, there are a number of excellent textbooks and shorter articles that are available. These include:

Gilbert, G. N., & Troitzsch, K. G. (2005). Simulation for the social scientist. New York and Maidenhead, U.K.: Open University Press.Find this resource:

Macal, C. M. (2016). Everything you need to know about agent-based modelling and simulation. Journal of Simulation, 10(2), 144–156Find this resource:

Macal, C. M., & North, M. J. (2010). Tutorial on agent-based modelling and simulation. Journal of Simulation, 4(3), 151–162Find this resource:

Railsback, S. F., & Grimm, V. (2012). Agent-based and individual-based modeling: A practical introduction. Princeton, NJ: Princeton University Press.Find this resource:

Wilensky, U., & Rand, W. (2015). An introduction to agent-based modeling: Modeling natural, social, and engineered complex systems with NETLogo. Cambridge, MA: MIT Press.Find this resource:

There is also a comprehensive book that covers a large and diverse number of geographical agent-based applications:

Heppenstall, A. J., Crooks, A. T., See, L. M., & Batty, M., (Eds.). (2012). Agent-based models of geographical systems. New York: Springer.Find this resource:

While the uptake of agent-based modeling in criminology is not broad, there are some particularly important contributions. These include:

Liu, L., & Eck, J. (2008). Artificial crime analysis systems: Using computer simulations and geographic information systems. Hershey, PA: Information Science Reference.Find this resource:

A collection of chapters discussing specific models as well as some broader aspects around the use of ABM to study crime:

Groff, E., & Mazerolle, L. (2008a). Simulated experiments and their potential role in criminology and criminal justice. Journal of Experimental Criminology, 4(3), 187–193.)Find this resource:

Simulated Experiments in Criminology and Criminal Justice. [Special Issue] (see

The two models discussed in detail in this entry were:

Birks, D., Townsley, M., & Stewart, A. (2012). Generative explanations of crime: Using simulation to test criminological theory. Criminology, 50(1), 221–254.Find this resource:

Malleson, N., Heppenstall, A., See, L., & Evans, A. (2013). Using an agent-based crime simulation to predict the effects of urban regeneration on individual household burglary risk. Environment and Planning B: Planning and Design, 40(3), 405–426.Find this resource:

Modeling human behavior is one of the most important, and most difficult, aspects to building agent-based models. For information about how to build complex behavior into agents using cognitive frameworks, the reader might refer to:

Balke, T., & Gilbert, N. (2014). How do agents make decisions? A survey. Journal of Artificial Societies and Social Simulation, 17(4), 13Find this resource:

Kennedy, W. G. (2012). Modelling human behaviour in agent-based models. In A. J. Heppenstall, A. T. Crooks, L. M. See, and M. Batty (Eds.). Agent-based models of geographical systems (pp. 167–179). Dordrecht, The Netherlands: Springer.Find this resource:

References

Andresen, M. A., Linning, S. J., & Malleson, N. (2017). Crime at places and spatial concentrations: Exploring the spatial stability of property crime in Vancouver BC, 2003–2013. Journal of Quantitative Criminology, 33, 1–21.Find this resource:

Arthur, C. (2010). Why Minority Report was spot on. The Guardian.Find this resource:

Axelrod, R. (1997). Advancing the art of simulation in the social sciences. In D. R. Conte, P. D. R. Hegselmann, & P. D. P. Terna, (Eds.), Simulating Social Phenomena (pp. 21–40). Lecture Notes in Economics and Mathematical Systems 456. Berlin: Springer.Find this resource:

Balke, T., & Gilbert, N. (2014). How do agents make decisions? A survey. Journal of Artificial Societies and Social Simulation, 17(4), 13.Find this resource:

Balzer, W. (2000). SMASS: A sequential multi-agent system for social simulation. In R. Suleiman, K. G. Troitzsch, & N. Gilbert, (Eds.), Tools and techniques for social science simulation (pp. 65–82). Heidelberg: Physica-Verlag HD.Find this resource:

Batty, M. (2005). Agents, cells, and cities: New representational models for simulating multiscale urban dynamics. Environment and Planning A, 37(8), 1373–1394.Find this resource:

Batty, M. (2008). Fifty years of urban modeling: Macro-statics to micro-dynamics. In S. Albeverio, D. Andrey, P. Giordano, & A. Vancheri, (Eds.), The dynamics of complex urban systems: An interdisciplinary approach (pp. 1–20). New York: Springer Physica-Verlag.Find this resource:

Batty, M. (2012). Building a science of cities. Cities, 29(1), S9–S16.Find this resource:

Benenson, I., Omer, I., & Hatna, E. (2002). Entity-based modeling of urban residential dynamics: The case of Yaffo, Tel Aviv. Environment and Planning B: Planning and Design, 29(4), 491–512.Find this resource:

Birks, D. J., Donkin, S., & Wellsmith, M. (2008). Synthesis over analysis: Towards an ontology for volume crime simulation. In L. Liu & J. Eck (Eds.), Artificial crime analysis systems: Using computer simulations and geographic information systems (pp. 160–192). Hershey, PA: Information Science Reference.Find this resource:

Birks, D., Townsley, M., & Stewart, A. (2012). Generative explanations of crime: Using simulation to test criminological theory. Criminology, 50(1), 221–254.Find this resource:

Birks, D., Townsley, M., and Stewart, A. (2013). Emergent regularities of interpersonal victimization: An agent-based investigation. Journal of Research in Crime and Delinquency, 51(1), 119–140.Find this resource:

Bonabeau, E. (2002). Agent based modeling: Methods and techniques for simulating human systems. Proceedings of the National Academy of Sciences, 99(suppl. 3), 7280–7287.Find this resource:

Brantingham, P. L., & Brantingham, P. J. (1981). Notes on the geometry of crime. In P. L. Brantingham & P. J. Brantingham (Eds.), Environmental criminology (pp. 27–54). Thousand Oaks, CA: SAGE.Find this resource:

Brantingham, P. L., & Brantingham, P. J. (1993). Nodes, paths and edges: Considerations on the complexity of crime and the physical environment. Journal of Environmental Psychology, 13(1), 3–28.Find this resource:

Brantingham, P., & Brantingham, P. J. (2004). Computer simulation as a tool for environmental criminologists. Security Journal, 17(1), 21–30.Find this resource:

Bratman, M. E., Israel, D. J., & Pollack, M. E. (1988). Plans and resource-bounded practical reasoning. Computational Intelligence, 4(3), 349–355.Find this resource:

Chainey, S., & Ratcliffe, J. (2005). GIS and crime mapping. Chichester, U.K.: John Wiley.Find this resource:

Cioffi-Revilla, C. (2014). Introduction to computational social science. Texts in Computer Science. London: Springer.Find this resource:

Clarke, R. V., & Cornish, D. B. (1985). Modeling offenders’ decisions: A framework for research and policy. Crime and Justice, 6, 147–185.Find this resource:

Cohen, L., & Felson, M. (1979). Social change and crime rate trends: A routine activity approach. American Sociological Review, 44(4), 588–608.Find this resource:

Crooks, A., Castle, C., & Batty, M. (2008). Key challenges in agent-based modelling for geo-spatial simulation. Computers, Environment and Urban Systems, 32(6), 417–430.Find this resource:

Crooks, A. T., & Heppenstall, A. J. (2012). Introduction to agent-based modelling. In A. T. Crooks & A. J. Heppenstall, (Eds.), Agent-based models for geographical systems (pp. 85–105). New York: Springer.Find this resource:

Dray, A., Mazerolle, L., Perez, P., & Ritter, A. (2008). Policing Australia’s heroin drought: Using an agent-based model to simulate alternative outcomes. Journal of Experimental Criminology, 4(3), 267–287.Find this resource:

Eck, J., & Weisburd, D. (1995). Crime places in crime theory. In J. Eck & D. Weisburd (Eds.), Crime and place (pp. 1–33). Monsey, NY: Criminal Justice Press.Find this resource:

Elffers, H., & van Baal, P. (2008). Realistic spatial backcloth is not that important in agent based simulation research: An illustration from simulating perceptual deterrence. In L. Liu & J. Eck, (Eds.), Artificial crime analysis systems: Using computer simulations and geographic information systems (pp. 19–34). Hershey, PA: Information Science Reference.Find this resource:

Epstein, J., & Axtell, R. (1996). Growing artificial societies: Social science from the bottom up. Washington, DC: Brookings Institution Press.Find this resource:

Epstein, J. M. (1999). Agent-based computational models and generative social science. Complexity, 4(5), 41–60.Find this resource:

Evans, A. J. (2012). A sketchbook for ethics in agent-based modelling. In Association of American Geographers (AAG) Annual Meeting, 23–26 February 2012, New York. Available online.Find this resource:

Gilbert, G. N., & Conte, R. (Eds.). (1995). Artificial societies: The computer simulation of social life. London: UCL.Find this resource:

Gilbert, G. N., & Troitzsch, K. G. (2005). Simulation for the social scientist. New York: Open University Press.Find this resource:

Gilbert, N. (2004). Agent-based social simulation: dealing with complexity. Centre for Research in Social Simulation.

Gilbert, N., & Doran, J. (1994). Simulating societies: Computer simulation of social phenomena. New York and London: Routledge.Find this resource:

Grimm, V., Berger, U., Bastiansen, F., Eliassen, S., Ginot, V., Giske, J., et al. (2006). A standard protocol for describing individual-based and agent-based models. Ecological Modelling, 198(1–2), 115–126.Find this resource:

Grimm, V., Berger, U., DeAngelis, D. L., Polhill, J. G., Giske, J., & Railsback, S. F. (2010). The ODD protocol: A review and first update. Ecological Modelling, 221(23), 2760–2768.Find this resource:

Groff, E. (2007a). Situating simulation to model human spatio-temporal interactions: An example using crime events. Transactions in GIS, 11(4), 507–530.Find this resource:

Groff, E. (2008). Characterizing the spatio-temporal aspects of routine activities and the geographic distribution of street robbery. In J. Eck & L. Liu (Eds.), Artificial crime analysis systems: Using computer simulations and geographic information systems (pp. 226–251). Hershey, PA: Information Science Reference.Find this resource:

Groff, E., & Mazerolle, L. (2008a). Simulated experiments and their potential role in criminology and criminal justice. Journal of Experimental Criminology, 4(3), 187–193.Find this resource:

Groff, E., & Mazerolle, L. (2008b). Simulated experiments in criminology and criminal justice [Special Issue]. Journal of Experimental Criminology, 4(3), 187–193.Find this resource:

Groff, E. R. (2007b). Simulation for theory testing and experimentation: An example using routine activity theory and street robbery. Journal of Quantitative Criminology, 23(2), 75–103.Find this resource:

Guerry, A. M. (1831). Essai Sur la Statistique Morale de la France. Paris: Chez Crochard.Find this resource:

Hayslett-McCall, K., Qui, F., Curtin, K. M., Chastain, B., Schubert, J., & Carver, V. (2008). The simulation of the journey to residential burglary. In L. Liu, & J. Eck (Eds.), Artificial crime analysis systems: using computer simulations and geographic information systems (pp. 281–299). Hershey, PA: Information Science Reference.Find this resource:

Heath, B., Hill, R., & Ciarallo, F. (2009). A survey of agent-based modeling practices (January 1998 to July 2008). Journal of Artificial Societies and Social Simulation, 12(4), 9.Find this resource:

Heppenstall, A., Malleson, N., & Crooks, A. (2016). “Space, the final frontier”: How good are agent-based models at simulating individuals and space in cities? Systems, 4(1), 9.Find this resource:

Heppenstall, A. J., Crooks, A. T., See, L. M., & Batty, M (Eds.). (2012). Agent-based models of geographical systems. Dordrecht, The Netherlands: Springer.Find this resource:

Johnson, S. D. (2004). The stability of space-time clusters of burglary. British Journal of Criminology, 44(1), 55–65.Find this resource:

Johnson, S. D., Bernasco, W., Bowers, K. J., Elffers, H., Ratcliffe, J., Rengert, G. F. et al. (2007). Near repeats: A cross national assessment of residential burglary. Journal of Quantitative Criminology, 23(3), 201–219.Find this resource:

Kennedy, W. G. (2012). Modelling human behaviour in agent-based models. In A. J. Heppenstall, A. T. Crooks, L. M. See, & M. Batty (Eds.), Agent-based models of geographical systems (pp. 167–179). Dordrecht, The Netherlands: Springer.Find this resource:

Kravari, K., & Bassiliades, N. (2015). A survey of agent platforms. Journal of Artificial Societies and Social Simulation, 18(1).Find this resource:

Liu, L., & Eck, J. (2008). Artificial crime analysis systems: Using computer simulations and geographic information systems. Hershey, PA: Information Science Reference.Find this resource:

Liu, L., Wang, X., Eck, J., & Liang, J. (2005). Simulating crime events and crime patterns in RA/CA model. In F. Wang (Ed.), Geographic Information Systems and Crime Analysis (pp. 197–213). Reading, PA: Idea Publishing.Find this resource:

Lum, K., & Isaac, W. (2016). To predict and serve? Significance, 13(5), 14–19.Find this resource:

Macal, C. M. (2016). Everything you need to know about agent-based modelling and simulation. Journal of Simulation, 10(2), 144–156.Find this resource:

Macal, C. M., & North, M. J. (2010). Tutorial on agent-based modelling and simulation. Journal of Simulation, 4(3), 151–162.Find this resource:

Malleson, N. (2010). Agent-Based Modelling of Burglary (PhD dissertation). School of Geography, University of Leeds, U.K.Find this resource:

Malleson, N. (2012). Using agent-based models to simulate crime. In In A. J. Heppenstall, A. T. Crooks, L. M. See, & M. Batty (Eds.), Agent-based models of geographical systems (pp. 411–434). Dordrecht, The Netherlands: Springer.Find this resource:

Malleson, N., Heppenstall, A., & See, L. (2010). Crime reduction through simulation: An agent-based model of burglary. Computers, Environment and Urban Systems, 34(3), 236–250.Find this resource:

Malleson, N., Heppenstall, A., See, L., & Evans, A. (2013). Using an agent-based crime simulation to predict the effects of urban regeneration on individual household burglary risk. Environment and Planning B: Planning and Design, 40(3), 405–426.Find this resource:

Moss, S., & Edmonds, B. (2005). Towards good social science. Journal of Artificial Societies and Social Simulation, 8(4).Find this resource:

North, M. J., Collier, N. T., Ozik, J., Tatara, E. R., Macal, C. M., Bragen, M. et al. (2013). Complex adaptive systems modeling with Repast Simphony. Complex Adaptive Systems Modeling, 1(1), 3.Find this resource:

O’Sullivan, D., & Haklay, M. (2000). Agent-based models and individualism: Is the world agent-based? Environment and Planning A, 32(8), 1409–1425.Find this resource:

Park, A. J., & Buckley, S. (2015). Three-dimensional agent-based model and simulation of a burglar’s target selection. Presented at European Intelligence and Security Informatics Conference (EISIC), 7–9 Sept. 2015. IEEE Xplore Digital Library.Find this resource:

Quetelet, L. A. J. (1842). Research on the propensity for crime at different ages. Cincinnati, OH: Anderson.Find this resource:

Railsback, S. F., & Grimm, V. (2012). Agent-based and individual-based modeling: A practical introduction. Princeton, CT: Princeton University Press.Find this resource:

Rengert, G. F. (1996). The geography of illegal drugs. Boulder, CO: Westview.Find this resource:

Sampson, R. J., Raudenbush, S. W., & Earls, F. (1997). Neighborhoods and violent crime: A multilevel study of collective efficacy. Science, 277(5328), 918–924.Find this resource:

Schelling, T. (1969). Models of segregation. The American Economic Review, 59(2), 488–493.Find this resource:

Schmidt, B. (2000). The modelling of human behaviour. Ghent, Belgium: Society for Computer Simulation International.Find this resource:

Shaw, C., & McKay, H. (1942). Juvenile delinquency and urban areas. Chicago: University of Chicago Press.Find this resource:

Weisburd, D., & Amram, S. (2014). The law of concentrations of crime at place: The case of Tel Aviv-Jaffa. Police Practice and Research, 15(2), 101–114.Find this resource:

Wilensky, U., & Rand, W. (2015). An introduction to agent-based modeling: Modeling natural, social, and engineered complex systems with NETLogo. Cambridge, MA: MIT Press.Find this resource:

Wooldridge, M. J. (2009). An introduction to multiagent systems. Chichester, U.K.: John Wiley.Find this resource: