AGOF: A GFlowNet-Guided 2-Opt Framework for Vehicle Routing Problems

Ni Zhang, Ling Pan, Zhiguang Cao, Hoong Chuin Lau

Keywords: Vehicle Routing Problem, GFlowNet, 2-opt

TL;DR: We propose a gflownet-guided 2-opt framework to solve the exiting opt solver's genelization problem

Abstract: The 2-Opt algorithm is a widely used classical search method in vehicle routing problems (VRPs). However, existing learning-based approaches designed for 2-Opt rely on autoregressive (AR) architectures, which suffer from limited generalization and high computational overhead. In this work, we propose the first non-autoregressive (NAR) framework for 2-Opt, which addresses the generalization and efficiency limitations of prior AR-based models by reducing the complexity of the input space, enhancing the robustness of the reward landscape, and eliminating the need for repeated inference during optimization. To enable effective training within this framework, we introduce A GFlowNet-guided 2-Opt Framework (AGOF), which leverages the reward–edge alignment capabilities of Generative Flow Network (GFlowNet) to provide accurate and generalizable edge evaluations for guiding 2-Opt swaps. Furthermore, we design Exploration beyond Local Optima (ELO) to inject perturbations into the optimization process, helping the model escape local optimal solutions. Extensive experiments demonstrate that AGOF not only outperforms existing GFlowNet- and 2-Opt-based methods but also has favorable generalization and computation efficiency.

Primary Area: optimization

Submission Number: 17892