Two-Stage Learned Decomposition for Scalable Routing on Multigraphs
This work introduces NodeEdge Policy Factorization (NEPF) to solve Vehicle Routing Problems (VRPs) on complex multigraphs. It splits routing into two stages: a node permutation stage and an edge selection stage. This decoupling allows scalable learning by separating high-level node decisions from low-level edge choices, leading to faster training and better performance.