Comment on paper: Position: Rethinking Post-Hoc Search-Based Neural Approaches for Solving Large-Scale Traveling Salesman Problems

In recent years, machine learning (ML) has emerged as a promising avenue for addressing optimization problems like the Travelling Salesman Problem (TSP). ML techniques, particularly those involving neural networks and reinforcement learning, have shown potential in learning heuristics and patterns that can guide the search for optimal routes more efficiently. By leveraging data, ML models can improve the quality of solutions and reduce computation time. One notable approach is the use of heat map-based search, where ML models generate heat maps that highlight promising regions of the solution space. These heat maps are then used to focus the search process, potentially enhancing the efficiency and effectiveness of finding optimal or near-optimal solutions [1]. Recently, the authors of paper [2] (referred to as SoftDist) discussed the neural approach and claimed: Our theoretical and experimental analysis raises doubts about the effectiveness of MLbased heat map generation. In support of this, we demonstrate that a simple baseline method can outperform complex ML approaches in heat map generation. Here, however, we show that the authors in SoftDist misconducted the experiments, leading to an unfair comparison and a flawed conclusion.