Efficient Online City-Scale Patrolling by Exploiting Offline Model-Based Coordination Policy

Abstract

With the increasing need of protecting public security, this paper studies the city-scale patrolling (CSP) problem, where hundreds of police officers are planned to patrol thousands of regions in a city. Given the stochastic nature with uncertain incident occurrence and travel time, the online CSP, where the patrolling policy should be determined sequentially, is of special interest. The online CSP aims at the omnipresence patrolling, which denotes that there are always police officers nearby that can respond timely when an incident occurs. Existing exact combinatorial optimization approaches are time-consuming and cannot scale to online CSP scenarios. On the other hand, within the dynamic CSP environments, existing decentralized multiagent coordination-based approaches always converge to the suboptimal solution without any performance guarantee. To achieve the omnipresence patrolling in a real-time fashion, a novel two-stage CSP framework is proposed. In the first stage, an offline model-based patrolling policy (OFFLINECSP) is designed, where the historical data is used to build the model and the linear programming (LP) technique is proposed for the coordination policy. Guided by the benchmark OFFLINECSP policy, an efficient online patrolling (ONLINECSP) policy is proposed in the second stage, where the police prefers to serve the critical incidents and to patrol hot-spot regions. The theoretical result shows that the competitive ratio of ONLINECSP can be guaranteed. Finally, extensive experiments on the synthetic and real datasets are conducted to validate the proposed framework. The results demonstrate that compared with existing benchmarks, the proposed two-stage CSP framework can not only maximize the incident service rate but also scale well to CSP in a real-time fashion.