Smart scheduling of dynamic job shop based on discrete event simulation and deep reinforcement learning

In the era of Industry 4.0, production scheduling as a critical part of manufacturing system should be smarter. Smart scheduling agent is required to be real-time autonomous and possess the ability to face unforeseen and disruptive events. However, traditional methods lack adaptability and intelligence. Hence, this paper is devoted to proposing a smart approach based on proximal policy optimization (PPO) to solve dynamic job shop scheduling problem with random job arrivals. The PPO scheduling agent is trained based on an integration framework of discrete event simulation and deep reinforcement learning. Copies of trained agent can be linked with each machine for distributed control. Meanwhile, state features, actions and rewards are designed for scheduling at each decision point. Reward scaling are applied to improve the convergence performance. The numerical experiments are conducted on cases with different production configurations. The results show that PPO method can realize on-line decision making and provide better solution than dispatch rules and heuristics. It can achieve a balance between time and quality. Moreover, the trained model could also maintain certain performance even in untrained scenarios.