Reinforcement Learning for Dual-Resource Constrained Scheduling

Abstract

This paper proposes using reinforcement learning to solve scheduling problems where two types of resources of limited availability must be allocated. The goal is to minimize the makespan of a dual-resource constrained flexible job shop scheduling problem. Efficient practical implementation is very valuable to industry, yet it is often only solved combining heuristics and expert knowledge. A framework for training a reinforcement learning agent to schedule diverse dual-resource constrained job shops is presented. Comparison with other state-of-theart approaches is done on both simpler and more complex instances that the ones used for training. Results show the agent produces competitive solutions for small instances that can outperform the implemented heuristic if given enough time. Other extensions are needed before real-world deployment, such as deadlines and constraining resources to work shifts.