Scheduling unrelated parallel machines with sequence-dependent family setups via deep reinforcement learning-driven rollout and adaptive batching

초록

Unrelated parallel machine scheduling problems (UPMSPs) with sequence-dependent family setups (SDFS) have attracted significant attention in various manufacturing industries. In contemporary manufacturing systems, frequent changes in customer demands and shop floor environments lead to variations in the numbers of jobs, machines, and families, which poses a great challenge in obtaining high-quality schedules within a short time limit. Although deep reinforcement learning (DRL) has emerged as a promising approach for UPMSPs with SDFS, it still faces an inherent trade-off between schedule quality and adaptability to such variations. In this paper, we propose a novel scheduling framework that utilizes the rollout algorithm with an adaptive batching strategy driven by a DRL agent to minimize the mean weighted tardiness of UPMSPs with SDFS. Specifically, compact state and continuous action representations are introduced to solve unseen scheduling problems with different numbers and characteristics of jobs, families, and machines. Furthermore, we employ an actor-driven rollout metaheuristic that aims to generate feasible schedules for action candidates, and a critic-based adaptive batching strategy is devised to effectively reduce the solution space by consolidating consecutive same-family jobs. Numerical experiments demonstrate that the proposed approach outperforms the DRL-based method, rule-based methods, and metaheuristics while requiring a short computation time. © 2026 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

키워드