Solving Adversarial Patrolling Games with Bounded Error

Patrolling games are partially observable games played by two players, the defender and the attacker. The defender aims for detecting intrusions into vulnerable targets by following randomized routes among them, the attacker strives to maximize the probability of a successful (undetected) intrusion. We show how to translate patrolling games into turn-based perfect information stochastic games with safety objectives so that optimal strategies in the perfect information games can be transferred back to patrolling games. We design, to the best of our knowledge, the first algorithm which can compute an ε-optimal strategy for the defender among all (history-dependent) strategies.