Distributed Deep Reinforcement Learning-based Dynamic Routing for Agent-based Manufacturing Execution Systems

In smart manufacturing, routing decisions for work-in-process (WIP) are critical factors affecting operational performance, including lead time, makespan, and congestion. Conventional approaches, such as shortest-path algorithms and rule-based heuristics, rely on predefined cost structures and thus have limitations in adapting to dynamic production environments. In this paper, we propose a distributed dynamic routing framework employing deep reinforcement learning (DRL) within an agent-based manufacturing execution system (MES). The routing problem is formulated as a multi-agent Markov decision process, where each WIP unit acts as an independent agent selecting its next movement based on locally observable states. A parameter-sharing deep Q-network (DQN) is used for scalable learning and decentralized execution, and the reward structure is optimized to minimize the average lead time. Simulation results demonstrate that the proposed DQN-based approach significantly outperforms Dijkstra-based routing and tabular Q-learning while maintaining low computational latency, demonstrating its applicability to real-time, large-scale manufacturing environments.