Core Concepts
The proposed ensemble method integrates a global policy that learns from the complete VRP graph and a local policy that focuses on transferrable local topological features, achieving better cross-distribution and cross-scale generalization performance compared to state-of-the-art methods.
Abstract
The paper presents a novel ensemble method, named Ensemble of Local and Global policies (ELG), for solving Vehicle Routing Problems (VRPs) with better generalization performance.
Key highlights:
The method consists of two base policies: a global policy that learns from the complete VRP graph, and a local policy that focuses on transferrable local topological features.
The local policy restricts the state and action space to the K nearest valid neighbor nodes of the current node, capturing the intrinsic characteristics of VRPs that are transferable across diverse problem instances.
The global and local policies are jointly trained to perform cooperatively and complementarily, with the global policy providing strong in-distribution learning capacity and the local policy contributing to better out-of-distribution generalization.
Extensive experiments on two well-known benchmarks, TSPLIB and CVRPLIB, demonstrate that the proposed ELG method significantly improves both cross-distribution and cross-scale generalization performance compared to state-of-the-art methods.
The ELG method even performs well on real-world VRP instances with several thousand nodes, while most existing construction methods can hardly solve such large-scale problems directly.
Stats
The paper does not provide specific numerical data to support the key logics. The results are presented in the form of performance gaps and runtime comparisons.