A Simulation-Based Comparison Between Ant Colony-Based Strategy and Multi-Agent Reinforcement Learning for Police Patrol

This work compares police patrol strategies in a simulation environment to discrete events that integrates preventive actions in hotspots and reactive response to calls. The study uses real operational data from an urban police unit and evaluates three approaches: random patrolling, Bayesian Ant-based Patrol Strategy (BAPS), and Multi-Agent Reinforcement Learning (MARL) with Dueling DQN architecture. The MARL was trained in independent scenarios and evaluated in a test scenario common to the other strategies, ensuring direct comparability. The results indicate that the MARL significantly reduced the idleness of hotspots, surpassing the references in this preventive dimension. On the other hand, BAPS performed better in reactive metrics, especially queue and average response time. The findings highlight the potential of MARL but indicate the need to better calibrate the reward function.