A multi-agent policy collaborative optimization algorithm with consensus constraints for automated control systems

Multi-agent reinforcement learning (MARL) enables cooperative decision-making in distributed automated control systems but may suffer from policy divergence and inconsistent behaviours. This paper proposes a Multi-Agent Policy Collaborative Optimisation (MAPCO) algorithm that integrates parameter, representation, and action consensus into a unified optimisation objective. A neighbourhood-based distributed optimisation strategy further supports collaborative policy learning with limited communication. Experiments on Cooperative Navigation and Distributed Energy Management demonstrate that MAPCO improves reward, success rate, consensus consistency, convergence efficiency, robustness, and scalability compared with representative MARL methods. These results validate MAPCO for reliable distributed automated control.