Knowledge-Embedded RL Framework for Capacitated Vehicle Routing Problems
Researchers proposed a unified reinforcement learning framework for generalized capacitated vehicle routing problems (CVRPs). The framework embeds explicit problem-solving knowledge inspired by Route-First Cluster-Second heuristics. It decomposes CVRPs into two subproblems: route-first and cluster-second. Dynamic programming solves the second subproblem, guiding an RL-based constructive solver for the first. This addresses limitations of end-to-end RL approaches that lack explicit knowledge, improving solution quality. The work appears on arXiv as 2605.14416.
Key facts
- CVRP is NP-hard with broad logistics applications
- Real-world CVRPs involve diverse objectives and constraints
- Proposed framework uses Route-First Cluster-Second heuristics
- Two-level knowledge embedding: decomposition and dynamic programming
- Dynamic programming results guide RL-based constructive solver
- Framework mitigates partial observability
- Paper published on arXiv with ID 2605.14416
- Framework aims to unify solutions for generalized CVRPs
Entities
Institutions
- arXiv