Abstract
We study the problem of multi-robot perimeter patrol in adversarial environments, under uncertainty of adversarial behavior. The robots patrol around a closed area using a nondeterministic patrol algorithm. The adversary’s choice of penetration point depends on the knowledge it obtained on the patrolling algorithm and its weakness points. Previous work investigated full knowledge and zero knowledge adversaries, and the impact of their knowledge on the optimal algorithm for the robots. However, realistically the knowledge obtained by the adversary is neither zero nor full, and therefore it will have uncertainty in its choice of penetration points. This paper considers these cases, and offers several approaches to bounding the level of uncertainty of the adversary, and its influence on the optimal patrol algorithm. We provide theoretical results that justify these approaches, and empirical results that show the performance of the derived algorithms used by simulated robots working against humans playing the role of the adversary is several different settings.