Cooperative multi-agent system for production control using reinforcement learning

Multi-agent systems can limit the control problem in complex production systems and solve them more efficiently. However, they often show local optimization tendencies. This paper presents a novel approach for a cooperative multi-agent system, which uses reinforcement learning and considers global key performance indicators. For this purpose, a central deep q-learning module transfers its knowledge to the cooperative order agents. The order agent's experience is stored in a replay memory for subsequent reinforcement learning. Interdependencies between the characteristics of nonlinear production systems and learning parameters are investigated and the performance is evaluated in comparison to conventional methods of production control.