Data-Driven Reinforcement Learning Tracking of MASs Under Injection Attack: A Controller-Dynamic-Linearization Approach

A novel reinforcement learning (RL)-based model-free adaptive control (MFAC) strategy is proposed to address the consensus tracking control issue for multiagent systems subjected to data injection attacks launched in the communication channel. By virtue of the dynamic linearization technique, equivalent dynamic linear data models for nonlinear systems and unknown linear ideal controllers are provided with the aim of determining the controller structure. Meanwhile, to reduce the effects of injection attacks, the fuzzy logic system is used to approximate the unknown nonlinear function by an online learning approach. Moreover, an RL MFAC method is proposed by employing the actor-critic structure to optimize the parameter estimation performance. Rigorous proofs are presented to ensure that the closed-loop systems are uniformly ultimately bounded even in the presence of injection attacks. The validity and superiority of the provided algorithm are further demonstrated through representative simulations.