Neural solvers excel in simple routing but falter on complex constraints. CaR introduces the first general framework using explicit learning-based feasibility refinement and joint training to generate diverse solutions for lightweight improvement. It outperforms SOTA solvers in feasibility, quality, and efficiency on hard constraints.
Key Points
- 1.Introduces Construct-and-Refine (CaR) for explicit constraint handling in neural routing solvers
- 2.Joint training yields diverse solutions enabling 10-step refinement vs prior 5k steps
- 3.First shared construction-improvement representation via unified encoder
- 4.Superior feasibility, quality, and speed over classical and neural SOTA on hard constraints
Impact Analysis
CaR bridges the gap for neural solvers in real-world constrained routing like logistics, reducing reliance on heavy post-processing. It promotes paradigm unification, aiding broader AI optimization adoption.
Technical Details
Employs joint construction training for feasibility-suited solutions and lightweight refinement. Introduces shared encoder for knowledge transfer across construction and improvement phases.