A Deep Reinforcement Learning Approach for Production Scheduling with the Use of Dispatch Rules

This work proposes a framework of a Deep Reinforcement Learning (DRL) scheduling agent that solves the production scheduling problem by deciding the dispatch rules to be used for tasks allocation to resources. For training, Deep Q-Learning is selected as the learning method and a Deep Q Network (DQN) is used for the core of the agent. A Discrete Event Simulation (DES) model represents the production system and the agent proposes actions in the form of dispatch rules. The agent was trained and tested in a production system of a bicycle industry upon various production scenarios. The DRL scheduling agent proposes production schedules with a 7% less makespan compared to traditional methods, such as shortest processing time (SPT) and longest processing time (LPT) dispatch rules.