TuneNSearch: A hybrid transfer learning and local search approach for solving vehicle routing problems
Year: 2026
Type: article
Abstract: This paper introduces TuneNSearch, a hybrid transfer learning and local search approach for solving a large range of vehicle routing problems (VRP). Recently, multi-task learning has gained much attention for solving various VRP variants. However, this adaptability often comes at the expense of the models’ performance. To address this challenge, we propose pre-training a model on the multi-depot VRP (MDVRP), which is then fine-tuned for each specific VRP variant. TuneNSearch employs a Transformer-based architecture, augmented with a residual edge-graph attention network to capture the impact of edge distances and residual connections between consecutive layers. Furthermore, we apply a two-stage process that integrates an efficient local search algorithm after inference, which yields substantial performance improvements with minimal computational overhead. Results show that TuneNSearch outperforms existing state-of-the-art models trained for each VRP variant, requiring only one-fifth of the total training epochs, thus enabling a more efficient and faster knowledge transfer between the tested variants. Notably, on CVRPLIB and TSPLIB instances, TuneNSearch consistently outperforms existing models in nearly all tested instances, significantly reducing the optimality gaps to 3% or less across all VRP variants considered. Our approach demonstrates great generalization capabilities, excelling across different tasks, distributions and problem sizes, addressing a long-standing gap in the literature.
Source: Computers & Operations Research
Institutions: University of Coimbra, University of Cambridge
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,570
OpenAlex ID: W7129433741