Neural Combinatorial Optimization for Real-World Routing

TL;DR

VRPs in real-world logistics involve asymmetric distances and travel times, which are poorly captured by synthetic Euclidean benchmarks. The paper introduces RRNCO, a real-world VRP dataset built from OSRM with 100 cities and a novel encoder–decoder model that uses contextual gating and Neural Adaptive Bias to integrate node and edge features efficiently. Through reinforcement-learning training, RRNCO achieves state-of-the-art performance among neural combinatorial optimization methods on ATSP, ACVRP, and ACVRPTW, while supporting fast, on-the-fly instance generation. By releasing both the dataset and the code, the work enables robust evaluation and practical deployment of learning-based routing in diverse urban environments.

Abstract

Vehicle Routing Problems (VRPs) are a class of NP-hard problems ubiquitous in several real-world logistics scenarios that pose significant challenges for optimization. Neural Combinatorial Optimization (NCO) has emerged as a promising alternative to classical approaches, as it can learn fast heuristics to solve VRPs. However, most research works in NCO for VRPs focus on simplified settings, which do not account for asymmetric distances and travel durations that cannot be derived by simple Euclidean distances and unrealistic data distributions, hindering real-world deployment. This work introduces RRNCO (Real Routing NCO) to bridge the gap of NCO between synthetic and real-world VRPs in the critical aspects of both data and modeling. First, we introduce a new, openly available dataset with real-world data containing a diverse dataset of locations, distances, and duration matrices from 100 cities, considering realistic settings with actual routing distances and durations obtained from Open Source Routing Machine (OSRM). Second, we propose a novel approach that efficiently processes both node and edge features through contextual gating, enabling the construction of more informed node embedding, and we finally incorporate an Adaptation Attention Free Module (AAFM) with neural adaptive bias mechanisms that effectively integrates not only distance matrices but also angular relationships between nodes, allowing our model to capture rich structural information. RRNCO achieves state-of-the-art results in real-world VRPs among NCO methods. We make our dataset and code publicly available at https://github.com/ai4co/real-routing-nco.

Figures (4)

• Figure 1: [Left] Most NCO works consider simplified Euclidean settings. [Right] Our work models real-world instances where durations and travel times can be asymmetric.
• Figure 2: Overview of our RRNCO real-world data generation and sampling framework. We generate a dataset of real-world cities with coordinates and respective distance and duration matrices obtained via OSRM. Then, we efficiently subsample instances as a set of coordinates and their matrices from the city map dataset with additional generated VRP features.
• Figure 3: Our proposed RRNCO model for real-world routing.
• Figure 4: Study of our proposed model with different initial contexts -- coordinates, distances, and our Contextual Gating (CG) -- as well as Neural Adaptive Bias (NAB). CG and NAB perform the best, particularly in out-of-distribution (OOD) cases.