NEPF: Scalable Neural Routing on Multigraphs via Two-Stage Decomposition
A new neural approach called Node-Edge Policy Factorization (NEPF) has been developed by researchers to tackle Vehicle Routing Problems (VRPs) on multigraphs, which feature parallel edges that offer different travel choices with various trade-offs, such as distance versus time. Unlike conventional neural VRP techniques that operate on simple or Euclidean graphs, NEPF enhances scalability by dividing the routing policy into two phases: node permutation and edge selection. This method utilizes a pre-encoding edge aggregation strategy and a non-autoregressive framework for the edge selection phase, coupled with hierarchical reinforcement learning for integrated training. Tests on six VRP variants indicate that NEPF either matches or surpasses the quality of leading solutions while significantly speeding up training. The findings are available on arXiv, reference 2605.05389.
Key facts
- NEPF splits routing policy into node permutation and edge selection stages.
- Method uses pre-encoding edge aggregation and non-autoregressive architecture.
- Hierarchical reinforcement learning trains stages jointly.
- Tested on six VRP variants.
- Matches or outperforms state-of-the-art solution quality.
- Training is significantly faster than existing methods.
- Addresses scalability issues in multigraph routing.
- Published on arXiv:2605.05389.
Entities
Institutions
- arXiv