Dynamic scheduling for flexible job shop with new job insertions by deep reinforcement learning

• Published in: Applied soft computing Vol. 91; p. 106208
• Main Author: Luo, Shu
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2020
• Subjects: Deep Q network; Deep reinforcement learning; Dispatching rules; Flexible job shop scheduling; New job insertion; Deep Q network; New job insertion; Deep reinforcement learning; Dispatching rules; Flexible job shop scheduling
• DOI: 10.1016/j.asoc.2020.106208
• ISSN: 1568-4946

In modern manufacturing industry, dynamic scheduling methods are urgently needed with the sharp increase of uncertainty and complexity in production process. To this end, this paper addresses the dynamic flexible job shop scheduling problem (DFJSP) under new job insertions aiming at minimizing the total tardiness. Without lose of generality, the DFJSP can be modeled as a Markov decision process (MDP) where an intelligent agent should successively determine which operation to process next and which machine to assign it on according to the production status of current decision point, making it particularly feasible to be solved by reinforcement learning (RL) methods. In order to cope with continuous production states and learn the most suitable action (i.e. dispatching rule) at each rescheduling point, a deep Q-network (DQN) is developed to address this problem. Six composite dispatching rules are proposed to simultaneously select an operation and assign it on a feasible machine every time an operation is completed or a new job arrives. Seven generic state features are extracted to represent the production status at a rescheduling point. By taking the continuous state features as input to the DQN, the state–action value (Q-value) of each dispatching rule can be obtained. The proposed DQN is trained using deep Q-learning (DQL) enhanced by two improvements namely double DQN and soft target weight update. Moreover, a “softmax” action selection policy is utilized in real implementation of the trained DQN so as to promote the rules with higher Q-values while maintaining the policy entropy. Numerical experiments are conducted on a large number of instances with different production configurations. The results have confirmed both the superiority and generality of DQN compared to each composite rule, other well-known dispatching rules as well as the stand Q-learning-based agent. •A deep Q-network (DQN) is proposed to select the appropriate dispatching rules.•Seven generic state features are extracted to represent the production status.•Six composite dispatching rules are designed to minimize the total tardiness.•The DQN is trained by deep reinforcement learning combined with two improvements.•Numerical experiments have verified the effectiveness and generality of the DQN.