中国机械工程 ›› 2023, Vol. 34 ›› Issue (02): 201-207.DOI: 10.3969/j.issn.1004-132X.2023.02.010

• 智能制造 • 上一篇    下一篇

集成强化学习算法的柔性作业车间调度问题研究

张凯;毕利;焦小刚   

  1. 宁夏大学信息工程学院,银川,750021
  • 出版日期:2023-01-25 发布日期:2023-02-16
  • 通讯作者: 毕利(通信作者),女,1968年生,教授。研究方向为数据挖掘、智能调度。发表论文45篇。E-mail:billy68@nxu.edu.cn。
  • 作者简介:张凯,男,1998年生,硕士研究生。研究方向为人工智能、智能调度。发表论文1篇。E-mail:zhangk0212@163.com。
  • 基金资助:
    国家自然科学基金(62266034);宁夏重点研发项目(2021BEE03020)

Research on Flexible Job-shop Scheduling Problems with Integrated Reinforcement Learning Algorithm

ZHANG Kai;BI Li;JIAO Xiaogang   

  1. School of Information Engineering,Ningxia University,Yinchuan,750021
  • Online:2023-01-25 Published:2023-02-16

摘要: 将柔性作业车间调度问题转化为马尔可夫决策过程,提出了集成5种深度Q网络(DQN)优化的算法D5QN。构建马尔可夫过程中,提取一组特征来表述状态,通过调度规则的组合设计出三组动作,通过直接和间接两种方式共同描述奖励。与基于规则、元启发式和其他强化学习算法的比较证明,所提方法可进一步缩短求解时间,并具有可行性和有效性。

关键词: 作业调度, 智能调度, 马尔可夫决策过程, 强化学习

Abstract: The flexible job-shop scheduling problems were transformed into a Markov decision process, and an algorithm D5QN integrated with 5 kinds of deep Q-network (DQN) optimizations was proposed. In the constructing of Markov process, a set of features was extracted to describe the states, and 3 sets of actions were designed by composite rules. The rewards were mapped by direct and indirect methods. The proposed algorithm was compared with the algorithms based on rules, meta-heuristic, and other reinforcement learning, which verifies the proposed algorithm may further decrease the calculating time, and have feasibility and effectiveness. 

Key words:  , job scheduling, intelligent dispatching, Markov decision process, reinforcement learning

中图分类号: