Reinforcement Learning-Based Secure Control for 2-D Markov Jump Systems With Hybrid Cyber Attacks

Abstract

This article addresses the secure control problem of 2-D Markov jump systems (MJSs) subject to hybrid cyber attacks consisting of denial-of-service (DoS) attacks and false data injection attacks. To explicitly capture the impact of DoS attacks, an augmented system is constructed via a state-compensation approach, and the communication blocking caused by DoS attacks is fully considered. Under a zero-sum game framework, the secure control problem can be solved by a coupled game algebraic Riccati equation (CGARE), and sufficient conditions are derived to guarantee the existence of the CGARE's solution. A reinforcement Q-learning algorithm is proposed to search the solution to the secure control problem of the 2-D MJSs, which guarantees the reliability and security of the 2-D MJSs in the presence of hybrid cyber attacks and eliminates the reliance of traditional control methods on accurate system models. Ultimately, the convergence of the proposed algorithm is sufficiently analyzed and the effectiveness of the algorithm is verified by simulation results.