Multi-agent deep reinforcement learning for integrated production and maintenance optimisation in shared manufacturing
将共享制造中的集成生产与维护规划建模为马尔可夫决策过程，提出多智能体深度强化学习框架，通过协调分配、多离散动作分布和注意力机制提升效率，实验表明共享制造在需求高、缺货成本高时优势明显。
Coordinating production and maintenance is essential for optimising manufacturing systems but remains challenging due to inherent trade-offs and uncertainties. Shared manufacturing offers a flexible solution by allowing companies to access shared resources, maintain production during maintenance, and enhance overall adaptability and productivity. However, this promising manufacturing model has received limited attention within integrated production and maintenance planning (IPMP). This study bridges this gap by framing the IPMP within shared manufacturing as a Markov decision process. To solve this problem, we propose a multi-agent deep reinforcement learning framework that incorporates three key components: (1) a coordination allocation mechanism with reward reshaping to guide feasible decisions; (2) a multi-discrete action distribution combined with eligibility masking to streamline the decision process; and (3) an attention mechanism paired with generalised advantage estimation to enhance learning efficiency. Computational experiments demonstrate the effectiveness of the proposed framework, highlighting the increasing advantages of shared manufacturing as lost sales costs and demand increase, and shared resource costs decrease. These results indicate that stakeholders should incentivize manufacturing resource suppliers to participate, thereby fostering competition and reducing resource costs. Manufacturers, especially those experiencing high lost sales costs and demand pressures, stand to benefit significantly from adopting this manufacturing model.