Leveraging Structural Constraints for Diffusion-based Neural TSP Solvers
About
Neural combinatorial optimization has recently achieved strong results on the Euclidean Traveling Salesman Problem (TSP) using generative models such as diffusion and consistency models. State-ofthe-art approaches like FT2T combine fast consistency-based prediction with gradient-based inference time refinement. However, gradient search often incurs significant computational overhead and may not align with the discrete structure of feasible solutions. We introduce Projected Consistency Inference (PCI), a plug-and-play, retraining-free alternative that replaces gradient refinement with structure-aware projections: PCI decodes valid Hamiltonian tours from the consistency model output and applies a lightweight local search (e.g., 2-opt). PCI achieves an average optimality gap (OG) of 0.17% on TSP with 500 cities, and 0.31% on TSP with 1000 cities, outperforming FT2T best settings (OG 0.22% and 0.36%, respectively) while reducing the inference time up to 30 to 40%. PCI also exhibits lower variance and memory usage, and can surpass classical heuristics such as LKH3 in rapid solution generation. Our results demonstrate that structure-aware inference time operations provide a practical and principled path for neural TSP solvers, complementing training time objectives.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Traveling Salesman Problem | Uniform-TSP1000 | Optimality Gap0.31 | 44 | |
| Traveling Salesperson Problem | TSPlib 100-10^4 cities | Optimality Gap0.83 | 14 | |
| Traveling Salesperson Problem | Euclidean 2D TSP 128 random instances 500 cities (test) | Optimality Gap (%)17 | 10 |