An Interactive Operation Agent Scheduling Method for Job Shop Based on Deep Reinforcement Learning (in Chinese)

Abstract

Due to the NP-hard property of job shop scheduling problem (JSSP), it is difficult to obtain a high-quality solution in a limited time, and the random disturbance of production environment further aggravates the complexity of JSSP dynamic scheduling. Based on deep reinforcement learning, a novel interactive operation agent (IOA) scheduling model framework is proposed. Through analysis of the constraint relationship between process route and processing equipment among operations, the processing processes in job shop are constructed as operation agents. The interaction mechanism between operation agents is designed, and each agent can interact with each other and update its own feature vector according to their relationship. Further, a deep neural network is constructed based on the operation characteristics and the earliest processing time to fit the action value function. As a result, the scheduling model can generate the scheduling strategy according to the system state and the characteristics of each operation agent. Double DQN algorithm is used to train IOA scheduling model, and the introduction of empirical playback mechanism effectively breaks the correlation between sequence training samples. The trained model can quickly generate high-quality scheduling scheme, and effectively execute rescheduling production strategy in case of machine failure. Experimental results show that the proposed IOA scheduling method is superior to greedy algorithm and heuristic scheduling rules, and has good robustness and generalization ability.

Ruiqi (Ricky) Chen

PhD Student in ECRG