Data-Driven Fault-Tolerant Reinforcement Learning Containment Control for Nonlinear Multiagent Systems

This article concentrates on the data-driven containment problem for a class of nonlinear discrete-time multiagent systems via reinforcement learning. A novel two-layer control architecture is designed. In the first layer, a reference model is introduced with which all signals of the multiagent systems will reach synchronization. On account of the critic-actor neural network architecture, an adaptive neural network controller with a multigradient recursive reinforcement learning algorithm and less learning parameters method is designed to tackle the tracking issues and actuator faults. Then in the distributed control layer, the virtual containment control input is developed via policy iteration with critic-actor neural networks such that the containment error will converge to a small neighborhood of the origin. Note that the proposed method makes the solution of optimal containment control problem independent of system dynamics and takes energy costs into consideration. Besides, the semiglobally uniformly ultimately bounded property of signals in the closed-loop system and the policy iteration convergence are guaranteed. Finally, some numerical illustrations are attached to consolidate the effectiveness of our proposed mechanism.