<ul><li>Researchers propose a Rescaling Graph Convolutional Network (RsGCN) to enhance generalization of neural TSP solvers for scalable problems and reduce training costs.</li><li>RsGCN incorporates a Rescaling Mechanism to focus on scale-dependent features related to nodes and edges, stabilizing graph message aggregation and maintaining numerical consistency.</li><li>Efficient training with a mixed-scale dataset and bidirectional loss is utilized in RsGCN, along with a post-search algorithm called Re2Opt for further optimization.</li><li>Experiments show that RsGCN achieves remarkable generalization and low training cost, outperforming neural competitors with fewer learnable parameters and training epochs.</li></ul>

RsGCN: Rescaling Enhances Generalization of GCNs for Solving Scalable Traveling Salesman Problems

Discover more