Towards Efficient Constraint Handling in Neural Solvers for Routing Problems

This paper introduces Construct-and-Refine (CaR), a novel, general, and efficient constraint-handling framework for neural routing solvers. While neural solvers excel in computational efficiency for simple routing, their performance with complex constraints remains a challenge. CaR addresses this by employing explicit learning-based feasibility refinement, enhancing their applicability to real-world routing problems. This innovation has significant potential for improving logistics, transportation, urban planning, and network optimization, where efficient solutions to complex routing scenarios are crucial for operational success.