Scheduling Reentrant FlowShops: Reinforcement Learning-guided Meta-Heuristics

Abstract

The reentrant flowshop scheduling problems (RFSP) are ubiquitous in high-tech industries such as semiconductor manufacturing and liquid crystal display (LCD) production. Given the complexity of RFSP, it is significant to improve the production efficiency using effective intelligent optimisation techniques. In this study, four meta-heuristics assisted by two reinforcement learning (RL) algorithms are proposed to minimise the maximum completion time (makespan) for RFSP. First, a mathematical model for RFSP is established. Second, four meta-heuristics are improved. The Nawaz–Enscore–Ham (NEH) heuristic is utilised for population initialisation. Based on the problem characteristics, we design six local search operators, which are integrated into the four meta-heuristics. Third, two RL algorithms, Q-learning and state–action-reward–state–action (SARSA), are employed to select the appropriate local search operator during iterations to enhance the convergence in a local space. Finally, the results of solving 72 instances indicate that the proposed algorithms perform effectively. The RL-guided local search can significantly enhance the overall performance of the four meta-heuristics. In particular, the artificial bee colony algorithm (ABC) combined with SARSA-guided local search yields the highest performance.