Improving nonlinear process control with model-free multi-agent reinforcement learning and proportional-integral-derivative fusion

This paper investigates the use of model-free multi-agent reinforcement learning (MARL) for controlling complex nonlinear systems without requiring an explicit process model. A centralised MAPPO (Multi-Agent Proximal Policy Optimisation) framework is proposed to coordinate multiple agents, each regulating one of the system outputs. The approach is applied to two distinct benchmark problems: a mode-redundant (MR) switched power converter and a simplified photobioreactor. In both cases, hybrid controllers are designed that combine classical Proportional-Integral-Derivative (PID) structures with learning-based policy corrections. Simulation results demonstrate that the MAPPO-PID architecture outperforms standalone PID and fuzzy logic controllers, particularly in tracking time-varying references. Quantitative evaluation confirms the robustness, adaptability, and scalability of the proposed model-free control strategy across diverse nonlinear process dynamics.