A Neural-Assisted Combinatorial Optimizer with Adaptive Neighborhood Search

This study proposes a Neural-Assisted Combinatorial Optimizer with Adaptive Neighborhood Search (NANS) to address the challenges faced by unmanned delivery vehicles (UDV) in urban logistics, including dynamic demand fluctuations, road safety factors, and capacity constraints. Building on the traditional Capacitated Vehicle Routing Problem (CVRP) model, a mathematical model incorporating safety coefficients and constraints was developed, referred to as CVRP with Safety-Based Routing Preference (CVRP-SRP). To solve this complex constraint problem, NANS combines deep learning with heuristic algorithms. Specifically, a Light Encoder and Heavy Decoder (LEHD) deep learning model is integrated with an Adaptive Large Neighborhood Search (ALNS) algorithm. Using the Random Re-Construct (RRC) mechanism, the LEHD model generates multiple high-quality initial solutions. Building on these initial solutions, the ALNS algorithm applies various destroy-repair operators with an adaptive selection mechanism and incorporates a simulated annealing strategy to explore the solution space and perform local optimization. Experimental results demonstrate that NANS outperforms other algorithms in terms of both solution quality and computational efficiency. This study provides valuable insights into the efficient, safe, and flexible scheduling of UDV in urban logistics.