A Hybrid Learning-Optimization Framework for Dynamic Vehicle Routing Using Real-World Postal Delivery Data

作者：K. Danach, Samir Haddad, Jomana Al-Haj Hassan, J. Sayah, Merhej Joseph, Kallab Chadi · 发表于：Middle East and North Africa Communications Conference · 年份：2026 · DOI：10.1109/menacomm69507.2026.11588305

The dynamic vehicle routing problem (DVRP) is a fundamental challenge in modern logistics, characterized by timedependent travel costs and operational constraints. This paper proposes a novel hybrid learning-optimization framework that integrates graph neural networks, attention-based reinforcement learning, and classical local search heuristics to address largescale, real-world routing scenarios. The proposed approach leverages data-driven representations to construct high-quality initial solutions, which are subsequently refined through optimization techniques to ensure feasibility and efficiency. Experiments conducted on a real-world postal delivery dataset demonstrate that the method achieves near-optimal performance with an average optimality gap of less than 1%, while reducing computational time by up to an order of magnitude compared to exact solvers. Furthermore, the model exhibits strong adaptability in dynamic environments, outperforming state-of-the-art learning-based approaches in terms of routing cost and delay under time-dependent conditions. Ablation studies confirm the contribution of each component, highlighting the effectiveness of the hybrid design. These results demonstrate that combining machine learning with optimization provides a scalable and robust solution for complex routing problems, with significant implications for real-world logistics and transportation systems.