This study presents a hybrid optimization framework that integrates a Parallel Adaptive Large Neighborhood Search (PALNS) mechanism with Variational Autoencoders (VAEs) to enhance solution quality and computational efficiency in vehicle routing problems. The results demonstrate significant reductions in routing cost and improved convergence compared to traditional ALNS and other metaheuristics.
Vehicle Routing Problems (VRPs) are widely known to be NP-hard and play a central role in modern logistics and supply chain operations. The increasing need for scalable, generalizable algorithms has led to the incorporation of deep learning models into classical optimization heuristics. This research emerges from the need to explore solution spaces more intelligently through latent representations learned by generative models, while ensuring computational efficiency through parallel search processes.
The article, accepted for publication in the academic journal Evolutionary Intelligence (Springer Nature), introduces a hybrid architecture that combines a VAE—trained to capture structural patterns in VRP instances—with a parallel implementation of Adaptive Large Neighborhood Search. The VAE learns latent route representations that guide destruction and repair operators, increasing the diversity and strategic quality of the search. In parallel, multiple ALNS threads operate simultaneously, exchanging elite solutions to prevent premature convergence and accelerate computational performance.
Experimental evaluations were conducted using benchmark VRP instances (50–200 customers) and real routing data from Lima, Peru. The proposed methodology achieved average cost reductions of 4–5% relative to standard ALNS, exhibiting steeper and more stable convergence trajectories. The incorporation of the VAE also enabled latent-space analysis, providing interpretability and facilitating sensitivity assessments related to latent dimensionality and route structure.
This work advances the field of combinatorial optimization by demonstrating that generative deep learning models can be effectively integrated into adaptive metaheuristics to enhance both exploration quality and computational efficiency. The addition of robust parallelization supports scalability to larger and more complex VRP instances. Overall, the findings indicate that this hybrid approach is a promising direction for the development of next-generation optimization algorithms in routing and logistics.
