Distributed Optimal Control of Nonlinear Differential Graphical Games based on Reinforcement Learning

This paper introduces continuous time nonlinear differential graphical games and proposes an online distributed optimal control algorithm to solve them. In differential graphical games، each agent error dynamics and performance index depend on its neighbors’ information. The proposed online distributed policy iteration algorithm solves the cooperative coupled Hamilton-Jacobi equations. In this algorithm which is based on reinforcement learning، each agent uses an actor-critic neural network structure where the weights of these neural networks are tuned synchronously. While all actor-critic networks are learning، closed loop stability and convergence to optimal control laws are guaranteed. Finally simulation results demonstrate the validity and performance of the proposed algorithm.