关键词: 动态作业车间调度, 马尔可夫过程, 图神经网络, 深度强化学习, 注意力机制

Abstract: For the dynamic job shop scheduling problem with emergency order insertion event, this paper creates a dynamic job shop environment model with the objective of minimizing the early and late completion time of all workpieces. The scheduling problem is converted into a Markov process and solved by a DDQN (double deep Q-network) deep reinforcement learning algorithm combined with a graph neural network. Firstly, the graph neural network is used to do feature extraction on the state analysis map of the workshop so as to avoid the problem of relying on artificial experience to design the state definition, the attention mechanism is introduced to enhance the ability of the reinforcement learning intelligences to acquire the state information, and six sets of rule scheduling are used as the decision space of the intelligences, and secondly, a brand-new reward method is defined to strengthen the ability of guiding the learning of the intelligences. Finally, the effectiveness and feasibility of the solution model constructed in this paper are illustrated by multi-group controlled experiments.

Key words: dynamic job shop scheduling, Markov process, graph neural network, deep reinforcement learning, attention mechanism