Improving adaptive large neighborhood search: an evaluation of parallel approaches with deep learning integration

This paper introduces a hybrid optimization framework that enhances vehicle routing by integrating Parallel Adaptive Large Neighborhood Search (PALNS) with deep generative modeling. Our method uses Variational Autoencoders (VAEs) to extract latent representations of routing patterns, which dynamically guide neighborhood selection during the search. Unlike conventional heuristics that rely on handcrafted rules, our approach learns from historical solution data to balance exploration and exploitation. We conduct extensive computational experiments on benchmark CVRP instances and real routing data, showing significant improvements in solution quality and steeper convergence curves compared to standalone ALNS and other metaheuristics. While there is modest runtime overhead, the latency is consistent and within practical bounds. The results also highlight interpretability of latent features and their contribution to dynamic search behavior. This work bridges machine learning and large-scale combinatorial optimization, with practical implications for logistics and supply chain applications.