Learning-Guided Rolling Horizon Optimization for Long-Horizon Flexible Job-Shop Scheduling

Furthermore, when evaluating the performance on 600 operations FJSP (10, 20, 30) in Table 1, we see that option (1) and (2), results in a longer solve time but an improved makespan from the architecture without attention. We also note that option (3) is strictly dominated by the performance of the architecture without attention. We note that the TNR-TPR tradeoff on the performance and solve time aligns with our theoretical analysis, as fixing something that should not have been (low TNR) harms the objective but helps the solve time, while failing to fix something that should have been (low TPR) harms the solve time and also indirectly harms the objective (under a fixed time limit). Due to the time benefit of the architecture without attention and the relatively competitive objective, we believe it makes sense to keep the simpler architecture without attention in the main paper.Figure 7: Ablation neural architecture: Attention among the overlapping and new operations. The architecture follows Figure 1, but introduces an additional cross attention among the overlapping and new operations before output the predicted probability for each overlapping operation.