Hierarchical multi-agent deep reinforcement learning for dynamic flexible job-shop scheduling with transportation
提出分层多智能体深度强化学习框架，将决策分解为三层，结合模仿学习策略，解决含运输机器人的动态柔性作业车间调度问题，在完工时间和鲁棒性上优于现有方法。
In real-world job-shop scheduling problems (JSPs), transport robots, or ‘transbots’, play a crucial role in optimising efficiency by autonomously sequencing pick-up and delivery jobs to machines. To tackle the exponential growth in state and action spaces, and the increased uncertainty from transbots when integrating transportation to variants of JSPs, a novel hierarchical multi-agent deep reinforcement learning framework is designed to address the dynamic flexible job-shop scheduling problem with transportation (DFJSP-T). The decision space is decomposed into three levels: the high-level agent prioritises a job to process, the mid-level agent arranges machines to handle the assigned job, and the low-level agent links transbots to the predefined job-machine sequence. Each agent operates with individual state and action spaces, sharing limited information, such as global job status and progress. The proposed framework is enhanced by an imitation learning strategy that on-the-fly exploits the best heuristic method from a set of popular heuristics for solving JSPs. Experiments on generated numerical cases and benchmark problem variants demonstrate that our proposed method achieves superior makespan and robustness compared to other learning-based and heuristic approaches. The results also show a fast convergence rate and consistent decision-making time, indicating promising scalability and efficiency for more complex scenarios.