GeoRouteNet: Geometry-Enhanced Non-Autoregressive Neural Solver for the Traveling Salesman Problem

GeoRouteNet is a newly proposed geometry-enhanced non-autoregressive neural solver for the Traveling Salesman Problem (TSP), which addresses limitations in existing NAR approaches by integrating centered node features, learnable radial distance basis functions, and distance-aware graph attention. It achieves a 0.32% optimality gap on 10,000 random TSP50 instances and a 1.26% gap on TSP100, significantly improving performance on stratified TSPLIB EUC_2D instances from 17.12% to 3.60%. This model's architecture and training innovations, including multi-candidate self-comparison reinforcement learning, offer practitioners a robust framework for efficiently solving combinatorial optimization problems with enhanced generalization and stability under diverse conditions.