深度强化学习求解柔性装配作业车间调度问题

Deep reinforcement learning for flexible assembly job shop scheduling problem

针对产品动态到达的柔性装配作业车间调度问题,以最小化总拖期为目标,构建了基于事件点的数学规划模型,该模型包含加工机器分配、加工工序排序、装配站分配和装配工序排序四个决策序列,并提出了一种基于多智能体的深度强化学习算法进行求解.首先,所提出的算法包含四个智能体分别对应四个决策序列,智能体之间采用价值分解网络(VDN)协作策略;然后,构建基于拖期的复合回报函数,提取生产系统指标作为全局特征,完善各智能体的调度动作;最后,设计了精英经验库,充分挖掘高回报样本的价值.案例结果表明所提出的方法在不同场景下都优于现有经典调度规则和元启发式算法挖掘的调度规则.

The flexible assembly job shop scheduling problem with dynamic products arrival was addressed,to minimize total tardiness.A mathematical programming model was proposed based on event points,which contains four decision-making sequences:processing machine assignment,processing operation sequence,assembly station assignment,and assembly operation sequence.This model was solved by deep reinforcement learning algorithm based multi-agent.Firstly,the proposed algorithm consisted of four agents corresponding to four decision sequences,and multi-agent adopted a value decomposition networks(VDN)based cooperative strategy.Secondly,the reward function with tardiness was designed,the digital features of production system were extracted as global features,and the scheduling actions of each agent were defined.Finally,an elite experience pool was designed to fully exploit the value of high return samples.The experimental results show that the proposed method is superior to both classical heuristic rules and meta-heuristic rules in different scenarios.