DRL-Based Vehicle Routing with Mixture-of-Experts for Generalization

Researchers have introduced R2E-IG, a model aimed at generalizing solutions for Vehicle Routing Problems (VRPs) through Deep Reinforcement Learning (DRL). Current DRL approaches, which rely on uniform distributions, struggle with real-world variations. R2E-IG enhances the policy network by dividing it into several modules, which are then adaptively reassembled using an instance-level gating mechanism. This design incorporates Residual Refined Experts (R2E) to improve expressiveness via residual refinement. The gating mechanism is responsible for learning representations that are aware of distribution, directing inputs to the appropriate modules. This paper, which emphasizes cross-distribution generalization, is available on arXiv (2605.26776).