Distributed reinforcement learning for optimal control of nonlinear processes

This work develops distributed reinforcement learning (DRL) algorithms to address the optimal control problem of nonlinear systems. Unlike centralized RL methods that suffer from scalability issues, the proposed DRL frameworks enable each agent to learn the local optimal control policy for its corresponding subsystem, optimizing the overall performance of the system while considering the control policies of other subsystems. Two DRL schemes are introduced: A sequential DRL algorithm, which updates local policies one by one to ensure convergence, and an iterative DRL algorithm, which updates all policies in parallel to reduce computation time. Compared to centralized approaches, the proposed methods significantly improve scalability and computational efficiency while ensuring closed‐loop stability. Theoretical analysis is provided to guarantee convergence and stability, and simulation results on chemical process examples demonstrate the effectiveness of the proposed methods.