A Multiagent Planning Approach for Cooperative Patrolling with Non-Stationary Adversaries

摘要

Multiagent patrolling is the problem faced by a set of agents that have to visit a set of sites to prevent or detect some threats or illegal actions. Although it is commonly assumed that patrollers share a common objective, the issue of cooperation between the patrollers has received little attention. Over the last years, the focus has been put on patrolling strategies to prevent a one-shot attack from an adversary. This adversary is usually assumed to be fully rational and to have full observability of the system. Most approaches are then based on game theory and consists in computing a best response strategy. Nonetheless, when patrolling frontiers, detecting illegal fishing or poaching; patrollers face multiple adversaries with limited observability and rationality. Moreover, adversaries can perform multiple illegal actions over time and space and may change their strategies as time passes. In this paper, we propose a multiagent planning approach that enables effective cooperation between a team of patrollers in uncertain environments. Patrolling agents are assumed to have partial observability of the system. Our approach allows the patrollers to learn a generic and stochastic model of the adversaries based on the history of observations. A wide variety of adversaries can thus be considered with strategies ranging from random behaviors to fully rational and informed behaviors. We show that the multiagent planning problem can be formalized by a non-stationary DEC- POMDP. In order to deal with the non-stationary, we introduce the notion of context. We then describe an evolutionary algorithm to compute patrolling strategies on-line, and we propose methods to improve the patrollers performance.