Simulation As A Tool For Police Planning Research Paper

This sample Simulation As A Tool For Police Planning Research Paper is published for educational and informational purposes only. If you need help writing your assignment, please use our research paper writing service and buy a paper on any topic at affordable price. Also check our tips on how to write a research paper, see the lists of criminal justice research paper topics, and browse research paper examples.

Crime relies, directly or indirectly, upon an array of factors, ranging from the levels of concentration of wealth to the physical organization of the urban center under consideration. Modeling the highly interconnected nature of this social system has recently attracted attention in computer science. As experiments in this domain cannot be performed without high risks, because they result on loss of human lives, simulation models have been chosen as supporting tools for this process. Multiagent systems (MAS) primarily study the behavior of autonomous and organized groups of software agents with the purpose of providing solutions to complex problems that could not be achieved by each individual agent alone. Multiagent-based simulation systems have been successfully adopted because the inherent characteristics of the agents (e.g., autonomy, sociability, and pro-activity) facilitate the construction of more dynamic models, thus contrasting with conventional computer simulation approaches.

Other different approaches, dubbed in general as bio-inspired approaches, have recently been investigated in conjunction with MAS for crime modeling. Collective intelligence as that demonstrated by swarms and the evolutionary approach are two of the most prominent concepts and have been used in the design of crime models. Here, an overview of these concepts from a decision support perspective is done in order to describe how they have been applied for the development of multiagent-based crime simulation (MACS). Two particularly relevant issues in the MACS context are discussed: model calibration and model evaluation.

Fundamentals

According to Russell and Norvig (1995), an agent is a physical or abstract entity that can be viewed as perceiving its environment through sensors and acting on the environment through actuators. Ferber (Ferber 1999) considers a multiagent system (MAS) as comprising (i) an environment; (ii) a set of passive and active objects (agents); (iii) an assembly of relations, which link objects to each other; and (iv) operations making it possible for the agents to perceive, produce, consume, transform, and manipulate objects. One can distinguish three levels of organization of agents (Rocher and Sherif 1972): micro-social, groups, and global societies. The micro-social level concerns the interactions between agents and the various forms they relate to one another. The level of groups and of societies refers to the dynamics of intermediate structures like organization and cities, respectively. Multiagent systems have been successfully adopted in conjunction with simulation models, which are generally referred to as multiagent-based simulation (MABS) systems. According to Gilbert and Conte (1995), MABS are especially appropriate when one has to deal with interdisciplinary problem domains. Such an approach, which is bottom-up in nature, is also appropriate for the study of social and urban problems, since social or urban environments are dynamic, nonlinear, and composed of a great number of variables and entities. The main objectives behind the construction of MABS systems are the following:

To test hypotheses related to the emergence of macro-level behavior from interactions occurring at micro-social levels
To build theories that can contribute to a better understanding of sociological, psychological, and ethological phenomena
To integrate partial theories coming from different disciplines (e.g., sociology, cognitive psychology, and ethology) in a common theoretical framework

A particular kind of simulation, called geosimulation, addresses an urban phenomena simulation model with a multiagent micro-social approach to simulate discrete, dynamic, and event-oriented systems (Benenson and Torrens 2004). In geosimulated models, simulated urban phenomena are considered a result of the collective dynamic interaction among animate and inanimate entities that compose the map representation. The Geographic Information System (GIS) is responsible for providing the “data ware” in geosimulations.

The study of agent self-organization and related concepts, such as emergence, is another important concept within micro-social MAS. The basic idea is that societies of agents demonstrate intelligent behavior at the collective level out of simple rules at the individual level. Moreover, these individual rules often do not explain the behavior that is attained at the collective level. Swarm intelligence is characterized (i) by strictly local communication, (ii) by the formation of emergent spatial–temporal structures, and (iii) by the agent’s making stochastic decisions based on the local information available. One of the branches of swarm intelligence is ant-colony optimization (ACO), proposed by Bonabeau et al. (1999). ACO is a meta-heuristic model for solving combinatorial problems that can typically be represented as graphs. ACO gets inspiration from other areas of science, in this case, biological sciences having the special feature of adapting well to dynamic settings.

In a nutshell, ACO works by allowing agents (ants) to explore a search space, but it requires these ants to leave feedback information about locations with good solutions on the space itself. Agents are then attracted by the feedback left in the environment – the larger the amount of information (pheromone), the more attractive the agents find the position in the environment. In order to avoid early convergence to local optima, the approach assumes that the information left is volatile and impermanent; if no other activity occurs, a piece of information left in the environment “expires” or disappears over a certain period of time. Although never explored as a model for criminal behavior, ACO’s characteristics appeared from the start to be an ideal fit to this purpose, as we will discuss later. Briefly, it allows for investigating variations of a learning process with or without a social factor.

Genetic algorithms (GAs) (Holland 1975) are general-purpose search and optimization algorithms that comply with the Darwinian natural selection principle and with some principles of population genetics to efficiently design (quasi-) optimal solutions to complicated computational and engineering problems. Such meta-heuristics maintain a population of chromosomes, which represent plausible solutions to the target problem and evolve over time through a process of competition and controlled variation. The more adapted an individual is to its environment (i.e., the solution is to the problem), the more likely such individual will be exploited for generating novel individuals. In order to distinguish between adapted and non-adapted individuals, a score function (known as fitness function) should be properly specified beforehand in a manner as to reflect the main restrictions imposed by the problem.

Micro-Social Level For Crime Simulation

Despite the existence of several works on MABS that investigate aggregated features of crime by means of macro-simulation, tools for police planning have essentially focused on the microsocial level. These latter follow a conceptual framework defined upon the following features: the environment into which the agent is inserted, the agents (their perceptions, productions, transformations, and objects manipulation), and the interaction between them.

The environment is a space, which generally has a volume (Ferber 1999). The agents are situated in the environment or interact with objects inserted into it. In MACS, the environment can be real or artificial. Typically, when the designer wants to reproduce the real environment, Geographic Information Systems (GIS) using digitalized maps of a geographic area for representing, for instance, streets (Groff 2008), are used. Artificial environments reproduce the main features of a geographic area via abstractions such as grids (Brantingham and Tita 2008) or graphs. Despite the claim of Elffers and Van Baal (2008) that the use of real representation in MACS is not so relevant as the quality of the model, there is a trend in this direction mainly because it facilitates evaluation by visual comparisons. Another way to model the environment is by means of cellular automata (CA) typically to represent static objects. CA is a discrete model of regular grids in which the state of a cell at time t is a function of the status of a finite number of neighborhood cells at time t-1. Examples of environments modeled as a CA can be found in Liu et al. (2005).

The related work on MACS has typically been based on the routine activity theory (RAT) (Cohen and Felson 1979), which states that in order for a criminal act to take place, three elements must coexist: a motivated offender; a suitable target, either an object or a person that can be attacked; and the absence of capable guardians in charge of preventive actions. Therefore, most of the micro-social crime models are based on these three kinds of agents: criminals, guardians, and targets.

On a micro-social level, the criminal agent tries to commit crimes or to move. Each criminal is endowed with a limited view of the environment, measured in terms of a radius following a predefined unit of measure. Criminals have one or more points of departure that we are going to call “gateways.” Such points of departure represent places where criminals are likely to start out, e.g., their residences, metro stations, and bus stops, before committing crimes (Wright and Decker 1994). Target selection is typically probabilistic, based on factors such as target vulnerability, distance between the criminal and the target, and the criminal’s experience. The decision whether to commit a crime or not is made based on the existence of guardians within the radius of the criminal’s sight.

The criminal behavior can be modeled with a learning component that exploits the agent’s experience as well as with information coming from other criminal agents. The success rate of individual agents can be computed as the ratio of the number of successful crimes to the overall number of crimes attempted in their lifetime, as in Furtado et al. (2008), or based on preferences of criminals computed from data mining and based on discrete choice theory (Xue and Brown 2006). Criminals form communities wherein hints are shared. Due to the interconnection of the communities, such hints could be relayed to other criminals in other communities, and the rate at which this happens depends directly on the topology of the network of communities.

Usually there is a set of guardians (police or not) available, each one associated with a target area. A guardian can have a route of length n, which is defined as a set, and each component of which is a triple composed by the target area, the interval of time the guardian remains at the target, and the daily period (patrol shift) the routes refer to. Guardians can demonstrate deterministic or stochastic behavior. A deterministic guardian will always move to the same target area and at the same pace predefined as an input parameter. Police guardians are modeled following real data, since these data are available and known by police institutions. However, finding good routes can be an important goal in order to understand the impact of police patrol on crime prevention or reduction. In Reis et al. (2006), a genetic algorithm system called GAPatrol is devoted to the specification of effective police patrol route strategies for coping with criminal activities happening in a given artificial urban environment, which, in turn, mimics a real demographic region of interest. The approach underlying GAPatrol allows for the automatic uncovering of hot spots and routes of surveillance, which, in real life, are usually discovered by hand with the help of statistical and/or specialized mapping techniques.

Notice that, while many simulation studies are aiming at understanding and analyzing the role of various assumptions about the behavior of the respective agents as a function of the ever-changing constellation of other agents and, hence, have a theoretical emphasis, we like to stress here the use of simulations as a means to study the resulting pattern of offender agents and victim agents as a function of various strategies that guardian agents, especially police agents, could have. This use of simulation modeling exploits the method for investigating various possible police strategies on the simulated offence pattern and, hence, has a more applied character.

The locations to be chosen by criminals are referred to as targets, which can be differentiated with respect to their mobility. Commercial/entertainment establishment such as drugstores, banks, gas stations, lottery houses, and malls are fixed, while mobile targets are, e.g., citizens in movement. In Brantingham and Tita (2008), citizen movement is modeled according to the Levy probability distribution, while Furtado et al. (2008) and Liang et al. (2001) have concentrated their study on crimes against property, hence modeling only fixed target.

Targets have a state of vulnerability that can be either active or inactive. A vulnerable target means that it is perceivable by a criminal. Otherwise, it would not take part in the set of high-priority choices of that criminal. In this case, each target must have a probability of being vulnerable, which can follow a distribution based on past real crime data for the associated target type or based on the preference of criminals (Xue and Brown 2006). In Liu et al. (2005), a tension factor was introduced in the model by measuring the impact of crime events on human beings. After a crime in a region, the tension increases, while the vulnerability decreases. The attractiveness of a target can vary depending on cost and reward factors related to the selection of the target such as the location, income, and race composition of the area based on census data. In Furtado et al. (2008), an exponential temporal distribution is used and varies on a daytime basis. For each period and type of target, a value for a configurable parameter, l, must be determined at the start of the simulation in order to define the pace of occurrence of crimes. For instance, in the evening, drugstore robberies may follow a distribution based on a given value for l; whereas, in daylight periods, the crime temporal distribution might shift, achieving values four times higher for l. At any simulation tick, at least one target is made vulnerable in accordance with the temporal distribution associated with its related type. The state of vulnerability is essential as a parameter to control the pace of crimes per type as happens in real life. However, one of the limitations of using values for input parameters from historical real-data analyses, e.g., the pace of crimes (the l parameter), is that the simulation model will not be capable of identifying a change in the pace of crime occurrences if that change occurs during the simulation time. This problem becomes more significant as the simulation time increases.

Modeling The Interaction Between Agents

Direct interaction means that agents communicate with each other by means of message exchange and/or because they are part of a community or society. In order to be part of a society, it is imperative for an individual to establish social links with other peers. Different forms of interaction among the same individuals, even considering small groups, may take place simultaneously and may vary at different paces through time. One usual means to represent and analyze the (evolving) social structure underlying an organization of individuals is by resorting to the concept of social networks. Roughly speaking, a social network alludes to any formal, graph-based structure where individuals are represented by nodes and the social relationships that unite them are represented by links (ties) between those nodes. The topology of a social network is an important issue to be considered in the analysis thereof, as it helps to determine the network’s usefulness (from the viewpoint of the individuals that participate in the network).

The social interaction and learning aspects that underlie criminal activities were investigated in Sutherland’s seminal work (Sutherland 1974) in which the differential association theory was proposed. This theory advocates that interaction with others who are delinquent increases the likelihood of someone becoming and remaining a delinquent. That is, peers can play a crucial role in the development of values and beliefs favorable to law violation. In this theory, Sutherland elaborates nine postulates, out of which two are particularly relevant from a perspective of direct agent interaction:

Criminal behavior is learnable and can be especially learned through the interactions one establishes with other persons, typically through a verbal communication process.
The main part of the learning of criminal behavior occurs within intimate personal groups.

Another important result coming from works investigating social network models within the context of criminology is that social networks are a natural way to explain the concentration of crimes per area. Crime data analyzed from different regions, and even countries, usually reflect the fact that there are huge spatial (but also temporal) variations in the crime rate between different cities and between different regions in a city. In this regard, Glaeser et al. (1996) show that less than 30 % of the spatial variation of crime (both interand intracity) can be explained by differences in local attributes. The remaining 70 % can be explained by social interactions, which means that the agents’ decisions about crime are somewhat positively correlated. The authors also show that the impact of social relations is greater in thefts, burglaries, assaults, and robberies (i.e., crimes against property) than in homicides. It is worth mentioning the work of Bosse et al. (2007), which created a model to simulate social learning of adolescence-limited criminal behavior, and the work of Furtado et al. (2008), which designed a model in which social networks are used to model criminal communication. In the latter, the authors showed that the goal of the crime model of generating a power-law spatial distribution of crimes was correlated to the communication aspect modeled via the social network.

Indirect interaction is modeled by means of objects or variables sharing, typically represented in the environment. Other kinds of nonintentional forms of communication that are sent by diffusion or propagation into the environment, like signals, are also used. Ferber (1999) alerts to the limitation that the lack of semantics of signals can provoke. Since the signal is propagated in the environment, all the agents living there can perceive it. A cry of a citizen can be perceived by a guardian as a call for help as well as used by a criminal as a discovery of a potential prey. One important feature of these signals is that their intensity decreases with the distance from the source and with time. The concepts of tension (Liu et al. 2005) and conductivity (Dray et al. 2008) previously mentioned are examples of indirect interaction.

An example of a hybrid approach that uses direct and indirect communication is that of Furtado et al. (2008). Here, communication between the agents was proposed from the concepts of ant-based optimization augmented with a social network. In this model, criminals prefer to commit crime in locations known to be vulnerable, with high payoff, etc. In other words, their choice considers their preference and knowledge about the crime points. The link here to ACO is that, according to this approach, ants always choose their next location in the environment (the place they move toward) biased by a mechanism (the pheromone marks) of indirect interaction. Another indirect communication strategy that ACO offers is that it includes concepts intrinsically related to the notion of “collective.” In ACO, ants perform their local search tasks without dictating the whole colony’s behavior, which, in turn, is recognized as an emerging result coming from all these local activities. Each criminal has three possible actions: commit a crime, not commit a crime, and move to a certain location. In order to reach a decision whether to commit a crime or not, criminals make use of a probabilistic approach which is adapted from the context of ant-based swarm systems (Dorigo and Stutzle 2004).

Key Issues

In MABS, the calibration of the model is a crucial step of the design process. These models are typically characterized by the existence of a lot of parameters, which together determine the general behavior of the model. The development and the parameter setting of MABS models can be long and tedious if there is no accurate and systematic manner to explore the parameter space.

Genetic algorithm has been studied as an alternative for parameter calibration in MABS. The basic idea is to consider the tuning process as an optimization problem. The optimization function for the MABS would be the distance between the artificial model and the real system. In MACS, the calibration of certain parameters in which the complexity is high deserves special attention, e.g., the place from where each criminal starts out to commit crimes. Examples of these initial locations, here called gateways, are bus stops, metro stations, and slums. Usually there is no real data or theoretical model to help one configure those gateways in a crime simulation model. Moreover, it is a combinatorial optimization problem, i.e., the problem of assigning criminals to gateways. More formally, let G = G_i ; i = 1; … ; Ng be the set of gateways and C =C_j ; j =; … ; N_c be the set of criminals under consideration. The goal is to allocate each C_j to a G_iin a way that a quality measure F, somehow related to the aim of the simulation model, is maximized. In this allocation process, any gateway can be assigned to a criminal and all criminals must be allotted to one, and only one, gateway. Besides, more than one C_jcan be assigned to a given G_i (i.e., we have not imposed any limit over the number of criminals assigned to a gateway). Since this assignment problem is combinatorial in nature, the number of feasible gateway configurations is an exponential function of the number of possible gateways.

As important as the calibration of guardians in gateways is the calibration of the number of criminals. There is no real data to help on that and even estimations are very tough to do. Also, this number can change during the simulation or remain constant. Typically in many models, the number of criminals is constant during a simulation. The absence of mechanisms that could implement variability in the number of criminals, such as arrests, can be justified when the ultimate goal is to find good police patrol routes. Considering constant the number of criminals means that crime reduction is only attained by preventing a potential criminal from acting. We could say that, by doing so, there is a preparation of the model to work in the worst of the scenarios, i.e., no way to reduce the number of potential criminals. Thus, the preventive police-planning goal is to define strategies to cover the urban space in a way that could prevent crimes from occurring; events that change the number of criminals are irrelevant.

Social sciences have struggled with this topic due to the difficulty to conduct experiments in controlled environments. MABS have emerged as a tool for social analysis in a way similar to natural controlled experiments. However, model evaluation is one of the biggest challenges of MABS and MACS in particular. A typology for validation must consider the following aspects. The first aspect regards constructing validity to account for the difference between the real world and the rendering of the simulated environment. The challenge here is to design a model representing an approximation that won’t be too detailed or similar to the real world because, in this case, the model loses its pragmatic value, and rather than testing theories, it only enumerates what happens under a specific and limited set of conditions. On the other hand, the model must fit the purpose for which it has been created without variables that would bring excessive and unnecessary complexity. The second aspect refers to the internal validity of the model, also called verification. Basically, it refers here to the reliability of the software in generating a determined result from the inputs and processing function for which it was designed. Minor bugs and ill-defined implementations can be responsible for results that falsify the experiments. The third aspect refers to the external validity, meaning how reliable generalizations of the model are for populations lager than the samples. Statistical conclusion validity is another important aspect to be considered in terms of MACS evaluation. Typically, MACS are stochastic and their variation and unpredictability pose problems in the establishment of statistical validation. The identification of regularity that deviates from chance is essential in that context. This must be done across simulation runs in order to be convincing. Finally, they describe the need for models to be evaluated as to their empirical validity via comparisons with real data. In geosimulation, for instance, visual comparisons between hot spots generated from simulations are plotted on maps for comparison with real-data hot spots. Several studies conducted in different countries and different contexts have shown that crime is not uniformly distributed in space and that some victims or targets have a much greater risk of victimization than others (Pease 1998). The temporal dimension also presents a nonuniform distribution with different types of crime having different rhythms (their periodicity) as well as time (their rate of occurrence). In Johnson et al. (2007), an analysis showing how crime clusters in space and time was provided. Following victimization at one’s home, those nearby experience an elevated risk of victimization, which decays as time elapses. Another strategy for validating simulation models is to find patterns in real data that indicate crime distribution instead of only relating to the exact numerical values. Doing so, it is possible to compare the results of the simulation with expected distribution of events.

Notice that the use of simulation for investigating possible police strategies is, to some extent, less hampered by the type of problems we have discussed in the present section. In such applications we simply vary the parameters governing the guardian (police) agents, in a way that they represent the strategies to be tested. Actual parameters about the police process (e.g., number of officers available at particular times, priorities) may be available from the police force for whom the simulation is run, thus limiting the variability and hence the number of simulations to be run.

Future Directions

Simulation of criminal activities in urban environments is an asset to decision-makers seeking to find preventive measures. Law enforcement authorities need to understand the behavior of criminals and their response in order to establish safety measures and policies. A conceptual framework for micro-social multiagent-based crime simulation was described involving concepts from Computer Science and AI in particular. Special attention was given to calibration and evaluation aspects, since they constitute open issues demanding further investigations and techniques.

Future investigations in the calibration field are in fact advance in terms of decision support systems, since some parameters like a police patrol route and criminal–gateway composition can shed light on non-understood aspects of preventive policing. In this context, such investigations are expected to provide satisfactory answers to questions like: How far from the optimal patrol routing strategies are those that are actually adopted by human police managers? How complex do such optimal patrolling routes need to be in terms of their total lengths and urban area coverage?

Bibliography:

Benenson I, Torrens PM (2004) Geosimulation: objectbased modeling of urban phenomena. Comput Environ Urban Syst 28(1/2):1–8
Bonabeau E, Dorigo M, Heraulaz G (1999) Swarm intelligence: from natural to artificial systems. Santa Fe Institute Studies in the Sciences of Complexity Series. Oxford Press
Bosse T, Gerritsen C, Treur J (2007) Cognitive and social simulation of criminal behavior: the intermittent explosive disorder case. AAMAS 58:367–374
Brantingham P, Brantingham P (1979) Environment, routine, and situation: toward a pattern theory of crime. In: Clark R, Felson M (eds) Routine activity and rational choice, vol 5. Transaction books, pp 259–294
Calvez B, Hutzler G (2005) Automatic tuning of agent-based models using genetic algorithms. Fourth international joint conference on autonomous agents & multi agent system, Netherlands
Cohen L, Felson M (1979) Social change and crime rate trends: a routine approach. Am Sociol Rev 44:588–608
Dorigo M, Stutzle T (2004) Ant colony optimization. The MIT Press, Cambridge
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: Liu L, Eck J (eds) Artificial crime analysis systems: using computer simulations and geographic information systems, pp 19–34
Ferber J (1999) Multi-agent systems: an introduction to distributed artificial intelligence. Addison-Wesley
Furtado V, Melo A, Coelho AL, Menezes R (2008) Simulating crime against properties using swarm intelligence and social networks. In: Liu L, Eck J (eds) Artificial crime analysis systems: using computer simulations and geographic information systems, pp 300–318
Gilbert N, Conte R (1995) Artificial societies: the computer simulation of social life. UCL Press, London
Glaeser E, Sacerdote B, Scheinkman J (1996) Crime and social interactions. Q J Econ 111(2):507–548, MIT Press
Groff E (2008) Characterizing the spatio-temporal aspects of routine activities and the geographic distribution of street robbery. In: Liu L, Eck J (eds) Artificial crime analysis systems: using computer simulations and geographic information systems, pp 226–251
Holland JH (1975) Adaptation in natural and artificial systems. University of Michigan Press
Johnson S, Bernasco W, Bowers K, Elfers H, Ratcliffe J, Rengert G, Townsley M (2007) Space-time patterns of risk: across national assessment of residential burglary victimization. J Quant Criminol 32(3):201–219
Liang J, Liu L, Eck J (2001) Simulating crimes and crime patterns using cellular automata and GIS. UCGIS student award papers 2001. Retrieved from http://www. ucgis.org/f2oppor.html
Liu L, Wang X, Eck J, Liang J (2005) Simulating crime events and crime patterns in a RA/CA Model. In: Wang F (ed) GIS and crime analysis. Idea Group, pp 197–213
Pease K (1998) Repeat victimization: taking stock. Crime detection and prevention series paper 90. Home Office, London
Reis D, Melo A, Coelho ALV, Furtado V (2006) GAPatrol: an evolutionary multiagent approach for the automatic definition of hotspots and patrol routes. In: Sichman JS, Coelho H, Oliveira S (eds) Proceedings of IBERAMIA/SBIA 2006, Lecture Notes in Artificial Intelligence (LNAI) 4140, pp 118–127
Rocher G, Sherif P (1972) A general introduction to sociology: a theoretical perspective. MacMillan Canada
Russell S, Norvig P (1995) Artificial intelligence: a modern approach. Prentice Hall Series in AI, New Jersey
Scott JP (2000) Social network analysis: a handbook. Sage
Snook B (2004) Individual differences in distance travelled by serial burglars. J Invest Psychol Offender Profiling 1:53–66
Sutherland E (1974) Principles of criminology, 4th edn. J. B. Lippincott, Philadelphia
Wright R, Decker S (1994) Choosing the target. Burglars on the job: street life and residential break-ins. Northeastern University Press, Boston
Xue Y, Brown D (2006) Spatial analysis with preference specification of latent decision makers for criminal event prediction. Decis Support Syst 41(3):560–573