Solving flexible job-shop problem considering skilled workers via multi-domain graph attention network
提出一种基于多域图注意力网络的调度方法，通过提取工序、机器和工人间的复杂依赖关系，实现考虑技能工人的柔性作业车间高质量调度，实验表明优于传统规则和现有深度强化学习方法。
The Flexible Job-Shop Scheduling Problem (FJSP) plays a crucial role in Industry 5.0 by enabling adaptive resource allocation and improving system resilience in manufacturing systems. However, FJSP considering skilled workers can be challenging due to the complex prioritisation and processing dependencies between operations, machines, and workers. Recently, learning priority dispatching rules (PDRs) using Deep Reinforcement Learning (DRL) has achieved great empirical performance for solving FJSP. To address the aforementioned challenges, this article proposes a novel scheduling method for FJSP considering skilled workers leveraging Multi-domain Graph Attention Networks (MDGAT) to extract state representation of complex dependencies and realise high-quality decision-making. First, a heterogeneous scheduling graph for the state of the Markov decision process is designed that integrates states about operation, machine, and worker nodes as well as dependencies between them. Furthermore, an MDGAT is designed that includes the operation representation modelling, the machine representation modelling, the worker representation modelling, and the representation fusion module. By computing node embeddings, semantic embeddings, and global embedding of multiple elements in heterogeneous scheduling graphs, MDGAT captures the complex latent relationships among operations, machines, and workers for supporting high-quality decision-making. Experimental results demonstrate that MDGATRL outperforms traditional priority dispatching rules-based methods and existing DRL-based methods.