Real-Time Scheduling Scheme based on Reinforcement Learning Considering Minimizing Setup Cost

Abstract

This study starts with the idea that the process of creating a Gantt Chart for schedule planning is similar to Tetris game with only a straight line. In Tetris games, the X axis is M machines and the Y axis is time. It is assumed that all types of orders can be worked without separation in all machines, but if the types of orders are different, setup cost will be incurred without delay. In this study, the game described above was named Gantris and the game environment was implemented. The AI-scheduling table through in-depth reinforcement learning compares the real-time scheduling table with the human-made game schedule. In the comparative study, the learning environment was studied in single order list learning environment and random order list learning environment. The two systems to be compared in this study are four machines (Machine)-two types of system (4M2T) and ten machines-six types of system (10M6T). As a performance indicator of the generated schedule, a weighted sum of setup cost, makespan and idle time in processing 100 orders were scheduled. As a result of the comparative study, in 4M2T system, regardless of the learning environment, the learned system generated schedule plan with better performance index than the experimenter. In the case of 10M6T system, the AI system generated a schedule of better performance indicators than the experimenter in a single learning environment, but showed a bad performance index than the experimenter in random learning environment. However, in comparing the number of job changes, the learning system showed better results than those of the 4M2T and 10M6T, showing excellent scheduling performance.