Reinforcement Learning for Cooperative Control of Multi‐Agent Systems With Switching Topologies

ABSTRACT This paper presents a resilient learning‐based adaptive sliding‐mode consensus control framework for nonlinear multi‐agent systems subject to switching communication topologies and actuator‐side cyberattacks. A distributed sliding‐mode leader‐follower tracking consensus protocol is first developed to ensure robust coordination under network constraints. To overcome the performance limitations of heuristic gain selection in conventional adaptive sliding‐mode control, an actor‐critic reinforcement learning scheme based on heuristic dynamic programming is embedded to tune controller parameters online without requiring explicit knowledge of the plant dynamics. The critic network approximates the Hamilton‐Jacobi‐Bellman (HJB) value function, enabling near‐optimal adaptive consensus performance, while the actor network synthesizes an optimal adaptive control policy that preserves robustness. Lyapunov‐based analyses establish asymptotic reachability, tracking consensus under average‐dwell‐time switching topologies, and uniform boundedness despite actuator cyberattacks. The proposed framework is validated through theoretical analysis, real‐time experimentation on the OPAL‐RT 4610XG platform, and hardware‐in‐the‐loop experiments using a leader‐follower Quanser QUBE‐Servo rotary inverted pendulum. Results demonstrate faster consensus convergence, actuator‐friendly control action, and improved robustness compared with prevalent adaptive and learning‐based controllers.