Optimization of job shop scheduling problem based on deep reinforcement learning

Abstract

Aiming at the optimization problem of minimizing the maximum completion time in job shop scheduling, a deep reinforcement learning optimization algorithm is proposed. First, a deep reinforcement learning scheduling environment is built based on the disjunctive graph model, and three channels of state characteristics are established. The action space consists of 20 designed combination scheduling rules. The reward function is designed based on the proportional relationship between the total work of the scheduled operation and the current maximum completion time. The deep convolutional neural network is used to construct action network and target network, and the state features are used as inputs to output the Q value of each action. Then, the action is selected by using the action validity exploration and exploitation strategy. Finally, the immediate reward is calculated and the scheduling environment is updated. Experiments are carried out using benchmark instances to verify the algorithm. The results show that it can balance solution quality and computation time effectively, and the trained agent has good generalization ability to the scheduling problem in the non-zero initial state.