Dynamic job shop scheduling in human-machine manufacturing systems: Deep reinforcement learning approach

Abstract

This thesis concerns job shop scheduling problem in human-machine manufacturing systems. A key aspect of this problem is to carefully consider the characteristics of human operators. For instance, every human has different work preferences and abilities, and these properties can be important information on job allocation. Despite its’ importance, few studies have investigated the characteristics of human operators in job-shop scheduling problem. One of the major reasons for lack of the research is that the characteristics of human operators are usually hidden in nature, which makes the mathematical formulation of the scheduling problem difficult. To tackle this challenge, this thesis proposes deep reinforcement learning approach to implicitly capture the unobservable characteristics of the human operators. Numerical experiments demonstrate the proposed model can consider task competence level and fatigue development of individual human operators which are hidden but implicitly contained in accumulated historical data. The result that the proposed method gives the shortest flowtime compared to classical dispatching rules supports usefulness of considering the characteristics human operators in a job shop scheduling problem.