Adapting attackers and defenders patrolling strategies: A reinforcement learning approach for Stackelberg security games

This paper presents a novel approach for adapting attackers and defenders preferred patrolling strategies using reinforcement learning (RL) based-on average rewards in Stackelberg security games. We propose a framework that combines three different paradigms: prior knowledge, imitation and temporal-difference method. The overall RL architecture involves two highest components: the Adaptive Primary Learning architecture and the Actor-critic architecture. In this work we consider that defenders and attackers conforms coalitions in the Stackelberg security game, these are reached by computing the Strong Lp-Stackelberg/Nash equilibrium. We present a numerical example that validates the proposed RL approach measuring the benefits for security resource allocation.