Using reinforcement learning to slove flexible job shop scheduling problems (FJSP) consists of two stages: feature extraction from environmental observations and then mapping these features to policies. The effectiveness of the policies relies heavily on the quality of feature extraction, yet research in this area is scarce. To address this, we propose a novel network structure, T-GNN (Transformer-Graph Neural Network), which combines transformer and graph neural networks to extract features from disjunctive graphs, emphasising critical areas. Additionally, Layer Normalisation is utilised to address distribution discrepancies in training and testing data for FJSP, helping to prevent model overfitting. An evaluation method is introduced to assess the effectiveness of the extracted features across multiple datasets. Experimental results show that our model outperforms traditional priority dispatching rules in both synthetic and classical datasets, highlighting the significant impact of the first stage's feature extraction on overall scheduling performance.
Zhao et al. (Mon,) studied this question.