Abstract: Dispatching rules are widely used in real jobshop,to select the best dispatching rules under the specific scenarios and improve the adaptability and optimization ability of dispatching rules under uncertain dynamic production process,an improved jobshop dynamic scheduling algorithm based on dispatching rules and Q-learning was proposed.The dynamic flexible job shop scheduling problem with new job insertions randomly was addressed aiming at minimizing the maximum delay time.The state space representation method and the reward mechanism,as well as the search strategy based on Boltzmann sampling function were elaborated under Q-learning framework for improving the ability to explore and use rules.Besides,to inherit the interpretability of dispatching rules and be able to process those randomly inserted jobs in real-time,the action set was constructed using several classical dispatching rules such as Shortest Processing Time (SPT) and Earliest Due Date (EDD),which obtained optimal dispatching rules under specific scenarios at specific time points by supporting the agent.Simulation results confirmed that the proposed dispatching ruler and Q-learning based algorithm could effectively handle randomly inserted jobs and achieve good scheduling performances.

Key words: dynamic scheduling, Q-learning algorithm, dispatching rules, job shop scheduling

摘要： 为了在特定的作业条件下找到最优调度规则,提高调度规则在不确定动态条件下的自适应、自寻优能力,提出一种调度规则与Q学习算法集成的作业车间动态调度算法。考虑车间中作业随机到达的动态情况,以最小化最大延迟时间为调度目标,在Q学习框架下设计了新的状态特征、奖励机制以及以Boltzmann采样函数为主体的搜索策略,提高了算法探索和利用规则的能力;以最短加工时间优先和最早交货期等经典调度规则构成动作集,继承了调度规则的可解释性,使智能体能实时处理随机到达的作业任务,通过持续学习和迭代更新获得不同作业场景下的最优调度规则。仿真研究和对比测试验证了所提算法的优越性。

关键词: 动态调度, Q学习算法, 调度规则, 作业车间调度