Reinforcement Learning Environment for Job Shop Scheduling Problems

Abstract

The industrial growth of the last decades created a need for intelligent and autonomous systems that can propose solutions to scheduling problems efficiently. The job shop scheduling problem (JSSP) is the most common formulation of these real-world scheduling problems and can be found in complex fields, such as transportation or industrial assemblies, where the ability to quickly adapt to unforeseen events is critical. Using the Markov decision process mathematical framework, this paper details a formulation of the JSSP as a reinforcement learning (RL) problem. The formulation is part of a proposal of a novel environment where RL agents can interact with JSSPs that is detailed on this paper, including a comprehensive explanation of the design process, the decisions that were made and the key lessons learnt. Considering the need for better scheduling approaches on modern manufacturing environments, the limitations that current techniques have and the major breakthroughs that are being made on the field of machine learning, the environment proposed on this paper intends to be a major contribution to the JSSP landscape, enabling academics from different areas to focus on the development of new algorithms and effortlessly test them on academic and real-world benchmarks.