RADAR：基于非对称距离表示的学习路由

Recent advancements in neural solvers have demonstrated impressive performance on vehicle routing problems (VRPs), yet most of these approaches rely on symmetric Euclidean distances, limiting their applicability to real-world scenarios. A significant challenge lies in encoding the relational features present in the asymmetric distance matrices of VRPs. Previous attempts to directly encode these matrices often resulted in poor compact embeddings and limited generalization at scale. RADAR is introduced as a scalable neural framework that enhances existing neural VRP solvers with the capability to manage asymmetric inputs. It addresses asymmetry from both static and dynamic perspectives, employing Singular Value Decomposition (SVD) on the asymmetric distance matrix to initialize compact and generalizable representations.