NEURO-METAHEURISTIC APPROACHE FOR ADAPTIVE VEHICLE ROUTING IN URBAN TRANSPORT NETWORKS

Abstract

This dissertation addresses the challenge of optimizing vehicle routes in dynamic urban environments, where traffic conditions, customer requests, and vehicle states change dur ing operations. It proposes a hybrid framework, NMFAVRK, that combines Graph Neural Networks (GNN) and a Genetic Algorithm (GA) to solve the Dynamic Vehicle Routing Problem with Time Windows (DVRPTW). The framework builds spatiotemporal customer representations to extract intelligent visit priorities used to initialize and improve the search, then reinforces them with local search and fast route repair when unexpected events such as congestion, breakdowns, or new requests occur. Experiments on the Solomon C102 bench mark show that the system achieved a total distance of 828.9 with a very small gap from the best known solution and outperformed a random GA baseline by a notable margin. It also responded to disruptions within fractions of a second while reducing fuel consumption under the green routing model. The results confirm that combining deep learning with metaheuristic search improves solution quality and operational flexibility.