Application of Reinforcement Learning with Recurrent Neural Networks for Optimal Scheduling of Flow-Shop Systems Under Uncertainty

This study presents a methodology for the application of an intelligent agent for optimal scheduling of flow-shop manufacturing systems subject to uncertainty in processing times and demands. The agent is trained through a Deep Reinforcement Learning (DRL) algorithm referred to as Deep Recurrent Q-Learning (DRQN). The novelty of this work lies in the use of Recurrent Neural Network (RNN) as the structure of the agent, never considered before for scheduling of chemical manufacturing plants. This network aims to identify correlations between consecutive events (time-series) which are useful for the decision-making process of the agent for solving flow-shop scheduling problems. A reward function is set to guide the agent to a) minimize the makespan of the process inside a horizon, b) satisfy the demands without overproducing products, and c) account for uncertainty in processing times. The results show that this modelling framework can produce an agent that is able to re-schedule operations online due to realization of uncertainty and without the need to solve additional (online) optimization problems.