Optimal Cooperative Control of Multi-Agent Systems Through Event-Triggered Model-Free Reinforcement Learning

Abstract

This paper addresses the optimal cooperative control problem for nonlinear multi-agent systems with completely unknown dynamics and proposes a learning control scheme based on the event-triggered mechanisms. The problem is reformulated as a multi-agent differential graphical game, and an off-policy integral reinforcement learning algorithm is introduced by deriving off-policy Bellman equations. To reduce the computational burden of the controller, an event-triggered mechanism is integrated into the adaptive learning process. To overcome the limitations of static triggering, the dynamic variable is introduced to utilize past triggering information. The theoretical proof demonstrates the asymptotic stability of the system and a numerical example validates the effectiveness of the proposed control scheme. Finally, in the case of the multiple manipulator system, a comparison of four control schemes shows that the proposed method not only ensures the system's control performance but also achieves a larger triggering interval, reducing the update frequency of the controller and saving communication bandwidth.