Relational Graph Attention-Based Deep Reinforcement Learning: An Application to Flexible Job Shop Scheduling with Sequence-Dependent Setup Times

Amirreza Farahani, M. A. H. van Elzakker, Laura Genga, Pavel Troubil, Remco M. Dijkman

Abstract: This paper tackles a manufacturing scheduling problem using an Edge Guided Relational Graph Attention-based Deep Reinforcement Learning approach. Unlike state-of-the-art approaches, the proposed method can deal with machine flexibility and sequence dependency of the setup times in the Job Shop Scheduling Problem. Furthermore, the proposed approach is size-agnostic. We evaluated our method against standard priority dispatching rules based on data that reflect a realistic scenario, designed on the basis of a practical case study at the Dassault Systèmes company. We used an industry-strength large neighborhood search based algorithm as benchmark. The results show that the proposed method outperforms the priority dispatching rules in terms of makespan, obtaining an average makespan difference with the best tested priority dispatching rules of 4.45% and 12.52%.