Abstract: Existing researches focus on constructing supply-demand matching models and developing solving algorithms for cloud manufacturing platforms,with insufficient attention to the impact of batch matching time horizon in uncertain environments on platform operations.Aiming at the complex scenario where capacity suppliers and demanders randomly arrive and may depart anytime in cloud manufacturing platforms,a Markov Decision Model (MDP) was established based on dynamic bipartite graphs and a Q-learning dynamic time horizon decision-making method utilizing state and action reshaping techniques was proposed.This method adaptively determines the matching time horizon based on aggregated information from platform orders and shared capacities,and generates stable matching solutions considering the preferences of suppliers and demanders.Numerical experiments demonstrate that the algorithm outperforms commonly used random-event-triggered and fixed matching time horizon methods in terms of comprehensive platform operational indicators across various scenarios and problem parameter changes.The experimental results provide management insights for the operation of supply-demand matching in cloud manufacturing platforms.

Key words: cloud manufacturing, shared platform, supply-demand matching, reinforcement learning, matching time horizon

摘要： 现有研究侧重于构建云制造平台供需匹配模型并开发求解算法,对批处理匹配时域长度在不确定环境下对云制造平台运营的影响关注不足。文章针对云制造平台产能供需双方随机到达并且可能随时离开的复杂情景,建立了基于动态二部图的马尔科夫决策模型（MDP）,并提出了基于状态和动作重塑技术的Q-learning动态时域匹配决策方法。该方法根据平台订单和共享产能的聚合信息,自适应地决策匹配时域长度,并产生考虑了供需双方偏好的稳定匹配方案。数值实验结果表明该算法在多种情景和问题参数变化下,在综合平台运营指标上优于常用的随机事件触发和固定匹配时域方法。实验结果为云制造平台供需匹配运营提供了管理启示。

关键词: 云制造, 共享制造, 供需匹配, 强化学习, 匹配时域