Reinforcement learning-based distributed adaptive consensus control for stochastic non-affine multi-agent systems subjected to switching topologies and hybrid uncertainties

This work focuses on the distributed adaptive consensus control design for stochastic multi-agent systems (MAS) characterized by hybrid uncertainties and switching topologies. All followers are subjected to matched and mismatched disturbances, stochastic nonlinearities, and non-affine structures, which complicate the controller design process. Firstly, the actor-critic framework is formulated to mitigate the internal uncertainties and resulting nonlinearities arising from stochastic disturbances. In addition, disturbance boundary estimators are developed based on the high power of errors to handle the adverse effects of multiple disturbances. Furthermore, augmented differential equations are formulated via auxiliary integration to introduce additional affine control input. Consequently, several innovative distributed controllers are developed, integrating reinforcement learning, disturbance boundary estimation, and dynamic surface control approach. By utilizing the invariant property of communication topology graph within the dwell time, the system stability under switching topologies is proved; The Lyapunov function analysis demonstrates that all signals within the MAS are semi-globally uniformly ultimately bounded, with the output of the followers converging to a compact neighborhood of the leader’s output. The effectiveness and superiority of the established controller is illustrated through numerical simulation examples.