Multi-objective flexible job-shop scheduling via graph attention network and reinforcement learning: Multi-objective flexible job-shop scheduling

In real-world production scheduling, it is crucial to quickly create a plan while also achieving various objectives. Consequently, addressing the multi-objective flexible job-shop scheduling problem (MOFJSP) is both complex and challenging. Previous methods utilizing meta-heuristic approaches have m...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:The Journal of supercomputing 2025, Vol.81 (1)
Hauptverfasser: Li, Yuanhe, Zhong, Wenjian, Wu, Yuanqing
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In real-world production scheduling, it is crucial to quickly create a plan while also achieving various objectives. Consequently, addressing the multi-objective flexible job-shop scheduling problem (MOFJSP) is both complex and challenging. Previous methods utilizing meta-heuristic approaches have made significant strides in approximating high-quality Pareto front. However, they have not adequately addressed the issue of prolonged computation times. This paper introduces an end-to-end approach to solving the MOFJSP that leverages graph attention networks (GATs) and reinforcement learning, which we term as multi-objective graph attention reinforcement learning scheduler. The GAT effectively captures the machine and operation features within heterogeneous graphs. We employ a weighted-sum method to decompose the problem into smaller optimization tasks, thereby balancing three scheduling objectives: minimizing makespan, maximum machine load, and total machine load. Experimental results demonstrate that the proposed method outperforms five commonly used multi-objective evolutionary algorithms on synthetic instances, with a more pronounced performance advantage observed in larger instances. Furthermore, results from solving public instances with model trained on the smallest synthetic instance ( 10 × 5 ) indicate that the proposed method can rapidly approximate the Pareto front, yielding high-quality solutions and effectively addressing unseen instances.
ISSN:0920-8542
1573-0484
DOI:10.1007/s11227-024-06741-2