A comparison of evaluation methods for police patrol district designs

Abstract

Police patrol district design presents a multi-objective optimization problem with two goals: minimizing workload variation between patrol districts and minimizing the response time for officers responding to calls for service. We evaluate three different methods for scoring district designs: a closed form probability based approach, a discrete-event simulation based on hypercube models for spatial queuing systems, and an agent-based simulation model. We find that all methods provide similar evaluations when service demand is low enough that cross-boundary support is infrequent. However, when the demand for service routinely exceeds the supply available within districts, only the agent-based simulation model accurately represents the resulting complexities and significantly changes the evaluation scores to reflect the behavior of the system.

Simulation optimization of police patrol districting plans using response surfaces

“…Instead of tick-by-tick evaluation of police cars’ state variables in the agent-based model, the discrete-event simulation in Zhang et al updates police cars’ status only when they get calls, arrive at scenes, leave scenes and return to patrol states. 25 The travel times are modeled using Euclidean distance and responding speed affected by overall traffic conditions at a different time of day. The generation of CFS incidents is the same as that used in the agent-based model.…”

“…The generation of CFS incidents is the same as that used in the agent-based model. The discrete-event simulation runs much faster than the agent-based model and provides a similar districting plan evaluation under low and median CFS incidents intensity levels 25 .…”

Police Districting Problem: Literature Review and Annotated Bibliography

“…Zhang et al (2013) [32] The focus of this research is the evaluation of three different methods for scoring police districting designs: a closed form probability based approach, a discrete-event simulation based on hypercube models for spatial queuing systems, and an agentbased simulation model. The scoring measures are evaluated on designs generated using the methodology presented in Zhang and Brown (2013) [29].…”

Localized Crime Prediction Methods

“…I would like extend this work to analyze the performance of prediction models that use a non-uniform grid. Moreover, researchers in [79] found that the temporal component can have a significant impact of three evaluation metrics used for scoring district designs. Motivated by their work, I would like to extend this analysis to study the impact of changing the prediction window (e.g., one day to one week) on evaluation metrics.…”