Harnessing heterogeneous graph neural networks for Dynamic Job-Shop Scheduling Problem solutions

Abstract

Manufacturing firms increasingly face the challenge of managing production complexity in the face of global competition and evolving customer demands. A critical component for maintaining efficiency and profitability in such an environment is the ability to solve the Job-Shop Scheduling Problem (JSSP) effectively. Traditional methods often fail in dynamic manufacturing settings, where unpredictability and the need for real-time adaptability can render static approaches obsolete. This study introduces a novel framework that uses Deep Reinforcement Learning (DRL) to address these limitations and navigate the intricacies of dynamic JSSP. By integrating heterogeneous graph neural networks (HGNNs) with DRL, we develop a model that not only captures the complex interconnections inherent in JSSP, but also dynamically adapts to the evolving nature of real-world manufacturing systems. Our proposed model is size-agnostic, meaning that it can deal with JSSPs with variable problem sizes. In addition, we introduce a new training method called Bootstrap Curriculum Learning (BCL) that enhances the model’s resilience and adaptability by training it in stages, progressively introducing the model to more challenging scheduling problem cases. The performance of the proposed model, evaluated against public benchmarks and synthetic datasets, shows superior performance and generalizability, offering a robust solution to dynamic scheduling challenges in smart manufacturing.