An Event-Based Reinforcement Learning Framework for Zero-Sum Differential Games in Multi-Agent Systems
Yilin Shang, Wenbo Zhang, Shan Xue, Liqi Wang, Weidong Zhang · 2025
This paper proposes an event-based reinforcement learning framework for zero-sum differential games in multi-agent systems (MASs), designed to improve computational and communication efficiency. First, the distributed optimal control problem is formulated as a zero-sum differential game using the minimax principle. The design includes a performance index function that accounts for both neighbor information and disturbance effects on each agent. Second, a novel event-triggering mechanism (ETM) is introduced to facilitate aperiodic control updates, enhancing flexibility while significantly reducing trigger frequency. Third, the coupled Hamilton–Jacobi–Isaacs equation (HJIE) is solved using a critic neural network (NN) in the reinforcement learning framework to derive Nash equilibrium. Finally, simulations are performed to validate the effectiveness of the proposed framework, demonstrating significant reductions in controller updates and improved resource efficiency.