Resilient reinforcement learning and robust output regulation under denial-of-service attacks

Abstract

In this paper, we have proposed a novel resilient reinforcement learning approach for solving robust optimal output regulation problems of a class of partially linear systems under both dynamic uncertainties and denial-of-service attacks. Fundamentally different from existing works on reinforcement learning, the proposed approach rigorously analyzes both the resilience of closed-loop systems against attacks and the robustness against dynamic uncertainties. Moreover, we have proposed an original successive approximation approach, named hybrid iteration, to learn the robust optimal control policy, that converges faster than value iteration, and is independent of an initial admissible controller. Simulation results demonstrate the efficacy of the proposed approach.