Neural Deconstruction Search for Vehicle Routing Problems

TL;DR

This paper introduces Neural Deconstruction Search (NDS), a two-phase framework for vehicle routing problems that deconstructs a current solution with a neural policy and rebuilds it using a greedy insertion. Trained via reinforcement learning, the deconstruction policy operates on an M-step sequential decision process, guided by seeds to promote exploration, while the reconstruction relies on simple, fast heuristics. The approach is instantiated for CVRP, VRPTW, and PCVRP and demonstrates competitive or superior performance compared to state-of-the-art OR methods under equal runtime, with strong scalability up to $N=2000$ customers and good generalization across instance distributions. The work highlights the practical impact of combining learned deconstruction with a lightweight greedy rebuild, achieving high-quality solutions with GPU-accelerated batch processing and offering a flexible framework adaptable to additional VRP variants and future improvements such as model distillation.

Abstract

Autoregressive construction approaches generate solutions to vehicle routing problems in a step-by-step fashion, leading to high-quality solutions that are nearing the performance achieved by handcrafted operations research techniques. In this work, we challenge the conventional paradigm of sequential solution construction and introduce an iterative search framework where solutions are instead deconstructed by a neural policy. Throughout the search, the neural policy collaborates with a simple greedy insertion algorithm to rebuild the deconstructed solutions. Our approach matches or surpasses the performance of state-of-the-art operations research methods across three challenging vehicle routing problems of various problem sizes.

Figures (11)

• Figure 1: Improving a solution via neural deconstruction.
• Figure 2: Improvement step of NDS.
• Figure 3: Comparison of NDS and SISRs performance across varying runtime limits.
• Figure 4: Scalability
• Figure 5: Validation performance throughout the training process.