4. An agent-based model of the urban environment
4.5 Agent movement
Many models of computational criminology base the movement patterns of agents upon the tenets of certain theories, such as RAT. They then compare observed patterns for such models against those produced by models of purely random movement. There remain nonetheless significant differences in how researchers have implemented models of movement, the timing of movement, the level of randomisation and the extent to which realistic movement patterns have been attempted. For example, as part of an agent’s routine, Groff (2008) assigns an agent 4 different activity nodes intended to represent a home location, a main activity such as work or school, and two other locations such as recreation spaces. No reason is provided for the use of four activity nodes, but the activity spaces are assigned depending upon the distributions of the population in terms of jobs and activities in Seattle, which she used as her example location. If an agent is unemployed and finds employment, or vice versa, their activity space can then change to adopt or drop an employment node. The pathways to and from these nodes are then spaces in which agents can be victims of street robbery. The agents move in a deterministic fashion between the nodes and replicate unique time patterns which have been assigned to them, giving them time to spend at work, in transit and at home and leisure locations. Groff’s focus is upon RAT and testing the premise that the more time spent away from home, the more likely an agent is to become a victim of crime. She therefore varies the percentage of time spent away from home and indeed provides support for routine activity theory.
Birks, Townsley and Stewart (2012) use 5 activity nodes per agent in their model but, like Groff, give no empirical reason as to why this number was chosen. An activity space is then formed as agents use transport nodes to move between activity nodes, and it is within this activity space that an awareness of potential targets is produced. In order to mimic reality, one of the activity nodes is designated as a home location, to which there is a high probability (0.8) that the agent will return once it has visited another node. This was chosen in order to ‘reflect the anchor-point-based navigation thought to be typical of human navigation (Golledge and Spector 1978)’ (ibid:233). If the agent does not return home, it choses another activity node and continues its journey. In his model of burglary in Leeds, Malleson (2010) effectively assigns his agents 3 or 4 activity nodes to move between: a home location, a work location (if
items. They navigate the road network, which is imported from a GIS map of the area, to an impressive level of complexity: routes are calculated to include areas which are suitable only for motor vehicles or pedestrians for example. Dijkstra’s shortest-path algorithm is used by agents to reach their destination in the quickest time possible. Malleson also employs a 24 hour architecture within which movements (or not in the case of sleeping) are conducted, all depending upon which motivation is highest at any one time. Wang (2009) provides his agents with 4 routine activity nodes, consisting of a bus stop, an office and two coffee shops, between which to move.
Thus far we have focussed upon the movement of ‘victim’ agents (or the burglar in Malleson’s model), but one class of agents – when included – which require their own movement rules within a model of urban crime are the ‘police’. Police patrol strategies have been the subject of a number of models in order to investigate optimal patrol strategies, either in abstract or in real environments using GIS layers. These models naturally vary in design depending upon the aims and resources of the authors, with Wang (2009) implementing 3 types of patrol: random; community policing (patrols shrunk to one area); and hotspot policing (concentrating
intensively in high crime areas). Dray et al (2008) again test the effectiveness of random patrols against hotspot patrols, while also incorporating problem-oriented patrolling as a third strategy. The random patrol strategy targets any area of the grid, hotspot policing focuses on areas with high risk values, while problem-oriented patrolling simulates police working in partnership with drug outreach workers to target areas of risk as assessed by both agents. This last strategy is particularly interesting due to the complex interaction between the agent classes allowing a truer representation of police movement on the ground. Jones, Brantingham and Chayes (2010) compare random patrols to ‘cops on the dots’ – their phrase for hotspot policing where their police agents move randomly but biased towards areas which are highly attractive to criminals. All of these models concur that hotspot patrolling is significantly more effective at reducing crime rates than random patrols, with Dray et al (2008) also finding that problem-oriented patrolling is even more effective than hotspot policing. Similarly to Johnson (2009), Fonoberova et al (2012) implemented a much more simplistic model on a grid system, with law enforcement officers having a limited field of vision and only arresting those within it, rather than adopting any specific patrolling patterns. Nevertheless, by naturally moving towards criminal agents, again this mimicked a crude hotspot strategy.
Within urban models of crime, ‘offender’ agents tend to move depending on the attractiveness of their target (either human or building agents), which is contingent upon their ability to sense the environment and their decision-making process, to which we now turn.
4.5.1 Awareness space
Developed from RAT, one of the central tenets of CPT (Brantingham and Brantingham 1993) is that offenders build up an ‘awareness space’ around their travel to and from regular
destinations, their ‘routine activities’. These awareness spaces are then those in which crimes are considered most likely to occur as offenders are familiar with them and the targets within them. Within the cognitive architecture of an agent, therefore, an awareness space can be developed in which targets may be located. Wang, Liu and Eck (2008) use cognitive maps to build their agents’ awareness of their environment during the model, with accumulated reward values of certain areas being updated the more time an agent spends within that area. Routine activity paths are developed during the model that then define the awareness spaces of the agents. Birks, Donkin and Wellsmith (2008) take a similar approach to this in their cops and robbers model, with grid cells representing spaces in which a numeric value of the agent’s awareness is attached. The more time the agent spends in a cell, whether moving or
committing an offence, the more their awareness of the cell and the opportunities within it grows. This is developed in later models (Birks, Townsley and Stewart 2012; 2013) where the effect of increased awareness spaces in areas of routine activity and travel between them is seen in the spatial clustering patterns of crimes committed by the agents. The experimental condition, in which the development of awareness space is activated, is compared to the control condition, in which awareness of all space is uniform and unchanging. Within these awareness spaces offender agents become conscious of potential targets (either human or building agents), but which target they choose to attempt an offence against depends upon how attractive that target is.
4.5.2 Target attractiveness
The attractiveness of a target to an offender agent will depend upon a number of criteria, including the level of guardianship in the vicinity of the target, the level of reward gained (or perceived) at the target, and the equation used to guide decision-making (discussed in detail in the next section). Guardianship was a concept used in a number of the models surveyed, and in many cases this was represented by the presence of other agents in the model. Groff (2007a; 2008) uses police agents in her models, with their proximity to a target reducing its attractiveness to offender agents. This formal guardianship is present in Dray et al (2008) who also incorporate another guardianship agent in the form of outreach workers whose role is to search for and assist drug users. Wang, Liu and Eck (2008) use place manager agents in their model to deter robbers from targeting a building, and their effectiveness in responding to a crime if it does occur makes it more or less attractive for an offender agent to return. Both Yonas et al (2011) and Malleson (2012) use a measure of collective efficacy amongst the
population as a guardianship device, with offender agents seeing areas of high collective efficacy as less attractive. Yonas et al look at the simulated witness response rate within an area – how likely the witness is to report a crime – and this then affects the likelihood of arrest and of an offender agent repeating their activity. Socio-demographic data from the census alongside other attributes are used by Malleson to create a community cohesion variable within his model, with areas of greater cohesion less attractive to offenders for fear of being caught.
The level of a target’s attractiveness may be uniform or may fluctuate depending upon the selection of variables introduced by the modeller. Furtado et al (2009) have fixed targets which fluctuate between being vulnerable or not vulnerable to offender agents, while Pitcher and Johnson (2011) have homes with varying levels of attractiveness in their model. The level of occupancy in the home varies depending upon the time of day in Malleson’s (2012) model, and therefore the level of guardianship and consequently attractiveness to the offender agent also varies. Wang, Liu and Eck (2008) and Groff (2007a; 2008) incorporate continuously updated risk levels at a location depending upon the presence and, in the former case, the efficacy of guardianship agents, which affect the attractiveness of a target to the offender. Wilhite and Allen (2008) model the number of criminals within an area (the target here being residence and criminal activity within the neighbourhood) affected by the levels of self-protection, neighbourhood protection and city protection. Varying these levels of protection across model runs varied the attractiveness of an area to the criminal agents.
Groff (2007a; 2008) has the offender agents in the models calculate the level of reward they would gain in order to determine the attractiveness of an agent. For example the wealth of potential targets for street robbery within the vicinity is assessed in order to determine which target will present the highest reward in monetary terms. This selection presumes a level of rational choice within the agent, and Groff combines this calculation with a level of error introduced to take into account other factors which affect decision making, but that are not explicitly modelled, to represent bounded rationality.
4.5.3 Agent decision-making
The BDI and PECS frameworks for agent decision-making, discussed in section 4.5, are complex architectures for multiple layered decisions, but much of the decision-making within the models surveyed here is based upon simpler equations calculated within the models using rational choice and bounded rationality. Perfect rational choice calculations presume that the agent will take into account all information available to it and make the decision which best optimises its outcomes by maximising utility. Bounded rationality however allows a level of
error which may be included in this decision-making, and allows for imperfect decisions to occur. This more accurately reflects real life decision-making, where humans rarely have complete information, or the ability to assess things optimally, and are often influenced by factors such as emotional and cognitive biases which can lead to sub-optimal decisions being made.
Of the 36 models analysed in table 4, 12 used perfect rational choice equations, 9 used bounded rationality, 3 used BDI and 8 used PECS (all written by Malleson and colleagues). The rest did not specify enough detail to determine which framework they used. The calculations for decision-making varied in complexity, but the use of bounded rationality using error terms is encouraging, with agent-based modelling being particularly suited to the use of such
stochasticity. Groff (2007a) makes the argument for the use of bounded rationality as it is well known in criminology that decisions are made by criminals based on a multitude of variables that they deem important, as well as others that are not necessarily under their control. The lack of a totally flexible and accurate incorporation of bounded rationality into the BDI architecture has been discussed above, but a level of stochasticity within an agent’s decision- making process is surely desirable in order to achieve a more realistic representation of this process.