Purity Law for Generalizable Neural TSP Solvers

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Liu, Wenzhao, Li, Haoran, Han, Congying, Zhang, Zicheng, Li, Anqi, Guo, Tiande
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915282310660096
author Liu, Wenzhao
Li, Haoran
Han, Congying
Zhang, Zicheng
Li, Anqi
Guo, Tiande
author_facet Liu, Wenzhao
Li, Haoran
Han, Congying
Zhang, Zicheng
Li, Anqi
Guo, Tiande
contents Achieving generalization in neural approaches across different scales and distributions remains a significant challenge for the Traveling Salesman Problem~(TSP). A key obstacle is that neural networks often fail to learn robust principles for identifying universal patterns and deriving optimal solutions from diverse instances. In this paper, we first uncover Purity Law (PuLa), a fundamental structural principle for optimal TSP solutions, defining that edge prevalence grows exponentially with the sparsity of surrounding vertices. Statistically validated across diverse instances, PuLa reveals a consistent bias toward local sparsity in global optima. Building on this insight, we propose Purity Policy Optimization~(PUPO), a novel training paradigm that explicitly aligns characteristics of neural solutions with PuLa during the solution construction process to enhance generalization. Extensive experiments demonstrate that PUPO can be seamlessly integrated with popular neural solvers, significantly enhancing their generalization performance without incurring additional computational overhead during inference.
format Preprint
id arxiv_https___arxiv_org_abs_2505_04558
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Purity Law for Generalizable Neural TSP Solvers
Liu, Wenzhao
Li, Haoran
Han, Congying
Zhang, Zicheng
Li, Anqi
Guo, Tiande
Machine Learning
Artificial Intelligence
Achieving generalization in neural approaches across different scales and distributions remains a significant challenge for the Traveling Salesman Problem~(TSP). A key obstacle is that neural networks often fail to learn robust principles for identifying universal patterns and deriving optimal solutions from diverse instances. In this paper, we first uncover Purity Law (PuLa), a fundamental structural principle for optimal TSP solutions, defining that edge prevalence grows exponentially with the sparsity of surrounding vertices. Statistically validated across diverse instances, PuLa reveals a consistent bias toward local sparsity in global optima. Building on this insight, we propose Purity Policy Optimization~(PUPO), a novel training paradigm that explicitly aligns characteristics of neural solutions with PuLa during the solution construction process to enhance generalization. Extensive experiments demonstrate that PUPO can be seamlessly integrated with popular neural solvers, significantly enhancing their generalization performance without incurring additional computational overhead during inference.
title Purity Law for Generalizable Neural TSP Solvers
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2505.04558