Solving the Job Shop Scheduling Problem with Reinforcement Learning: A Statistical Analysis

Abstract

This paper focuses on the application of modern reinforcement learning techniques to solve the job shop scheduling problem where the objective is to minimize the makespan. It starts by discussing the proposed system, which is based on an original architecture that proposes a novel approach based on the cooperation of different components, from the problem that needs to be solved to the job shop training environment or the reinforcement learning algorithm. Then, a methodical computational study is presented. The computational study is divided into two phases of performance analysis, namely of efficiency and effectiveness. In the efficiency analysis the execution times until a solution is obtained by the studied approaches are compared. In the efficiency analysis the quality of the solutions proposed by the chosen methods of optimization are evaluated. Both phases included a meticulous statistical inference analysis to validate the obtained results. A detailed discussion of the collected results is also included in this paper. Based on the statistical evidence that was found, it culminates in the conclusion that the proposed approach examined and evaluated in this paper is most likely one of the most efficient to date to solve job shop scheduling problems.