Graph Neural Network-Based Task Offloading and Resource Allocation for Scalable Vehicular Networks

Abstract

Intelligent vehicles require extensive data processing to enhance safety and improve driver comfort. With limited onboard computing resources, these vehicles often offload tasks to nearby vehicles or servers for auxiliary processing to meet real-time response requirements. However, the complexity and highly dynamic nature of the vehicular environment render the design of effective offloading strategies. While existing approaches can adapt to changes in environmental parameters within vehicular networks, they are fundamentally limited by their inability to process variable-dimensional environmental information and make decisions that scale with network size. Traditional methods typically rely on fixed-size input representations and static computational frameworks, which are inherently unsuitable for the dynamic and scalable nature of real-world vehicular networks that require adaptive responses to varying network sizes. As a result, existing alternatives lack feasibility to highly dynamic real-world vehicle networks that require adaptive responses to varying network sizes. To alleviate this limitation, we develop an original approach to address the task offloading and resource allocation problem with a scalable size, via a framework based on a graph neural network (GNN). Leveraging its neighbour aggregation mechanism, GNN effectively adapts to varying-scale topologies in dynamic vehicular networks, ensuring robust performance regardless of network size. To evaluate our proposed approach, we conducted extensive simulations to analyse its performance. The experimental results demonstrate that our method provides a more scalable and real-time capable solution, surpassing existing approaches by seamlessly handling dynamic network size variations.