Online learning for patrolling robots against active adversarial attackers

Abstract

We study the online route planning problem for patrolling robots, to assign them to optimal routes to patrol in a large crime-prone area. To model the actively engaging, intelligent, and adversarial opponents, we use the Stackelberg Security Game between the patrolling robots and the attackers. We leverage a graph-based bandit algorithm [16] with adaptive adjustment of the reward for the robots in this game to perplex the best response attackers and gradually succeed over them. Our graph bandits can outperform other stochastic bandit algorithms [10] when a simulated annealing-based scheduling mechanism is incorporated to adjust the balance between exploration and exploitation. Hence our method can successfully assign a small group of patrolling robots to cover a large number of routes.