Learning to Dispatch for Flexible Job Shop Scheduling Based on Deep Reinforcement Learning via Graph Gated Channel Transformation
Dainlin Huang, Hong Zhao, Lijun Zhang, Kangping Chen
Abstract
In addressing the Flexible Job Shop Scheduling Problem (FJSP), deep reinforcement learning eliminates the need for mathematical modeling of the problem, requiring only interaction with the real environment to learn effective strategies. Using disjunctive graphs as the state representation has proven to be a particularly effective method. Additionally, attention mechanisms enable rapid focus on relevant features. However, due to the unique structure of attention mechanisms, current methods fail to provide effective strategies after changes in scale. To resolve this issue, we propose an end-to-end deep reinforcement learning framework for the FJSP. Initially, we introduce a lightweight attention model, the Graph Gated Channel Transformation (GGCT), to identify the characteristics of the workpieces being scheduled at the current decision-making moment, while suppressing redundant features. Subsequently, to address the inability to provide effective strategies after changes in scale, we modify the expression of disjunctive graph features, channeling global features into different channels to capture relevant information at the current moment effectively. Comparative analysis on generated and classical datasets shows our model reduces the average makespan significantly, from 8.243% to 7.037% and from 10.08% to 8.69%, respectively.