Hierarchical Aggregation Deconstruction Search for Vehicle Routing Problems

Chuanbo Hua, Federico Berto, André Hottung, Nayeli Gast Zepeda, Yining Ma, Paula Wong-Chung, Changhyun Kwon, Cathy Wu, Kevin Tierney, Jinkyoo Park

Keywords: neural combinatorial optimization, vehicle routing problem, reinforcement learning, improvement method

TL;DR: We propose HADES, a neural improvement framework for VRPs that introduces anisometric, hierarchical positional encodings to better capture solution structure, enable state-of-the-art performance.

Abstract: Recent progress in neural combinatorial optimization has shown promise for vehicle routing problems (VRPs). Iterative improvement frameworks address the limitations of pure construction policies, which often struggle with exploration and large-scale performance, but they remain constrained by solution encodings that ignore the hierarchical structure of routes. We introduce Hierarchical Aggregation Deconstruction Search (HADES), a neural improvement method with distance-aware, anisometric positional encodings tailored to routing solutions. HADES incorporates two complementary components: an in-route positional encoding, which captures the circular and non-uniform ordering of nodes within tours, and a cross-route encoding, which represents route membership and structural relations across tours. This hierarchical design provides solution representations better aligned with the anisometric and head-tail connected nature of VRPs, leading to more effective deconstruction. Extensive experiments across multiple VRP variants demonstrate that our model consistently advances the state of the art, with particularly strong gains on large-scale benchmarks. We will make our source code publicly available to foster future research.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 24233