Learning-enabled Flexible Job-shop Scheduling for Scalable Smart Manufacturing

TL;DR

This work tackles the scale generalization challenge in DRL-based scheduling for flexible job-shop problems with transportation constraints (FJSPT) in smart manufacturing. It introduces Heterogeneous Graph Scheduler (HGS), a graph-based DRL framework comprising a heterogeneous graph, a structure-aware encoder, and a three-stage decoder to generate end-to-end operation-machine-vehicle decisions. The method demonstrates strong makespan performance and robust scale generalization, outperforming traditional dispatching rules, meta-heuristics, and existing DRL approaches on both small-scale and unseen large-scale instances, including those with automated guided vehicles. The findings suggest that locally encoded relational knowledge and a stage-wise decoding strategy enable effective, size-agnostic scheduling in dynamic, AGV-driven manufacturing environments, with practical implications for scalable deployment. $C_{max}$ minimization, network attention mechanisms, and dynamic edge updates are central to the reported gains, highlighting a path toward scalable, real-time MIS scheduling in smart factories.

Abstract

In smart manufacturing systems (SMSs), flexible job-shop scheduling with transportation constraints (FJSPT) is essential to optimize solutions for maximizing productivity, considering production flexibility based on automated guided vehicles (AGVs). Recent developments in deep reinforcement learning (DRL)-based methods for FJSPT have encountered a scale generalization challenge. These methods underperform when applied to environment at scales different from their training set, resulting in low-quality solutions. To address this, we introduce a novel graph-based DRL method, named the Heterogeneous Graph Scheduler (HGS). Our method leverages locally extracted relational knowledge among operations, machines, and vehicle nodes for scheduling, with a graph-structured decision-making framework that reduces encoding complexity and enhances scale generalization. Our performance evaluation, conducted with benchmark datasets, reveals that the proposed method outperforms traditional dispatching rules, meta-heuristics, and existing DRL-based approaches in terms of makespan performance, even on large-scale instances that have not been experienced during training.

Figures (6)

• Figure 1: Disjunctive graph for FJSP.
• Figure 2: Heterogeneous graph scheduler architecture.
• Figure 3: Heterogeneous encoder architecture. This figure shows the sub-encoding process of the given example graph. For simplicity, we only illustrate $\text{HMHA}_{ij}$ block between operation node $O_{ij}$ and its neighboring nodes in the heterogeneous multi-head attention layer.
• Figure 4: Three-stage decoder architecture.
• Figure 5: Training curve on 10$\times$6$\times$6 instances.