Agent-based simulation for predictive modeling : a study of economically motivated crime in Charlotte, NC

This thesis investigates the use of approach combining agent-based simulation and geographical information systems for predictive modeling. Specifically, it looks at a case study in the field of crime and tests the hypothesis that agent-based simulation models are a valid means of forecasting criminal behavior. The model supposes that using data mining procedures on past records can gamer the hidden target preferences of criminal agents. These discovered preferences can be molded into a set of behavioral rules for computer agents. Within the scope of an agent-based simulation, criminal agents then use these rules to mimic the behavior of the real criminals. The end result is a predictive model for criminal behavior based on historical data.

This project spans feature selection, clustering of records, behavioral rule determination, simulation implementation and the testing of forecasted results. After data normalization and organization, feature selection is performed to harvest a subset of salient predictor variables. These features are then used to cluster the data set and establish the number of criminal agents in the system as well as their individual preferences. A cluster specific salience weighting method (CSSW) is utilized in conjunction with Wards hierarchical clustering algorithm to derive preferences for each agent in the scope of the simulation. A multivariate likelihood function is then f01med for each agent as well as the null case. Prior probabilities of crime and the distribution of incidents across the agent set are then combined with the likelihood functions using a Bayesian model to obtain the probability of agent j committing a crime in geographic area x. After the calculation of these probabilities, agents are left to freely roam a digital replica of a real city.

In a simulation implemented using ESRI ArcGIS and Visual Basic, each agent begins at a random location and evaluates the surrounding locations, moving at each iteration to the adjacent location that maximizes his utility. While each agent adheres to his own simplistic rule set the system as a whole has the potential to exhibit very complex results. The simulation utilizes Geographical Information System (GIS) maps as the environmental foundation. The GIS map stores information on hundreds of predictor variables based on data collected by the U.S. Census. The rules for each agent iteratively refer to this set of information when processing each agent's next movement and action. The agent acts based on his historic preferences and the environmental factors embedded in the GIS maps.

If the simulation can accurately forecast the behavior of criminals then police may be able to use the prediction results to take preventative measures. Police have the potential to estimate what locations, time periods and premises are most likely to be targeted. After police are alerted to what areas will be targeted and when, they can patrol or increase security measures accordingly. This project is a template that simulates a single type of crime in a single city. It serves as a small piece of the bigger puzzle, but it is the first step at assessing the idea that computers can geographically model human behavior and consequently gives us insight into a criminal population's future behavior.

Note: Abstract extracted from PDF file via OCR.