Abstract: To solve the dynamic adjustment problem of key parameters of flexible job-shop solving algorithm,a deep self-learning tabu search algorithm (DSLTS) based on deep reinforcement learning is proposed.The algorithm takes Tabu Search (TS) as the basic optimization method,and uses double deep Q network (DDQN) to intelligently adjust the key parameters of TS algorithm.Firstly,the self-learning model of DSLTS algorithm is analyzed and established.LSTM network is used to fit multiple TS algorithm feature vectors,and the results are input into DDQN network for learning iteration.Secondly,the state feature vector and reward function of reinforcement learning under TS algorithm are designed.Finally,the effectiveness and performance of other common FJSP solving algorithms and DSLTS algorithm in solving FJSP problems are compared,and the effectiveness of the proposed model and method is verified.

Key words: shop scheduling, deep reinforcement learning, self-learning, tabu search algorithm

摘要： 为解决柔性作业车间调度求解算法关键参数动态调整问题,提出一种基于深度强化学习的自学习禁忌搜索算法(DSLTS)。该算法以禁忌搜索算法(Tabu Search,TS)为基础优化方法,并采用双层深度Q网络(Double Deep Q Network,DDQN)智能调整TS算法关键参数。首先,分析并建立DSLTS算法中的自学习模型,利用长短期记忆网络（Long Short-Term Memory,LSTM）网络拟合多条TS算法特征向量,将结果输入DDQN网络中进行学习迭代。接下来,设计了TS算法环境下强化学习的状态特征向量和奖励函数。最后,比较其它求解FJSP问题常见算法和DSLTS算法在求解FJSP问题时的求解效果和性能,验证所提模型和方法的有效性。

关键词: 柔性车间调度问题, 深度强化学习, 自学习算法, 禁忌搜索算法