RRNCO: Towards Real-World Routing with Neural Combinatorial Optimization

Jiwoo Son, Zhikai Zhao, Federico Berto, Chuanbo Hua, Zhiguang Cao, Changhyun Kwon, Jinkyoo Park

Keywords: Neural Combinatorial Optimization, Combinatorial Optimization, Vehicle Routing Problem, Reinforcement Learning

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, existing research works in NCO for VRPs learn from simplified, symmetric Euclidean settings, failing to handle the asymmetric distances and travel durations inherent to real-world road networks. This critical sim-to-real gap severely hinders their practical deployment. To address this fundamental limitation, we introduce RRNCO, a novel NCO architecture with two key innovations for handling real-world routing complexity. First, we propose an Adaptive Node Embedding (ANE) approach that fuses coordinate information with distance features through learned contextual gating. Unlike existing methods relying solely on spatial coordinates or requiring full distance matrix processing, our approach efficiently captures both local geometric structure and global routing constraints through probability-weighted distance sampling that prioritizes nearby nodes while preserving asymmetric relationships. Second, we introduce Neural Adaptive Bias (NAB), the first mechanism to jointly model asymmetric distance and duration matrices within a deep neural routing framework. NAB's gating-based architecture learns to dynamically integrate distance, duration, and directional angles into a unified contextual bias that guides the Adaptation Attention Free Module (AAFM). Together, these innovations enable RRNCO to explicitly capture real-world routing asymmetries where costs from location A to B differ from B to A due to traffic patterns, road directionality, and temporal dynamics. We validate our method on a newly constructed dataset featuring real-world asymmetric distance and duration matrices from 100 diverse cities. Experiments demonstrate that RRNCO achieves state-of-the-art performance among NCO methods on realistic VRPs. We release our dataset and code to advance research in practical neural combinatorial optimization.

Submission Number: 6