AI Optimizes Multi-Hop Relay for Urban V2X Networks.

Ensuring reliable and low-latency communication for Connected and Automated Vehicles (CAVs) in dense urban environments is critical for smart mobility, but challenges like limited Road-Side Unit (RSU) density and non-line-of-sight conditions often hinder stable single-hop connectivity. Multi-hop relay communication can extend coverage, but real-time relay selection under dynamic constraints is computationally demanding. Traditional optimal solutions, such as Mixed-Integer Linear Programming (MILP), suffer from high computational complexity that scales poorly with network density, making them impractical for real-time applications. To overcome this, a new Learning-to-Optimize (L2O) framework has been developed, leveraging Graph Neural Networks (GNNs). This framework models vehicular communication states as attributed graphs, with CAVs and RSUs as nodes and radio links enriched with propagation-aware features. An offline MILP oracle provides optimal supervision for training an edge-aware Graph Isomorphism Network (GINE, which then approximates these optimal decisions with near-constant inference latency. Experiments using large-scale urban datasets demonstrate that this GNN-based approach achieves connectivity levels similar to the MILP oracle, but with orders of magnitude faster execution. This innovation allows for cost-effective enhancement of urban V2X connectivity by utilizing existing vehicular assets and supporting scalable, real-time NR-V2X operations in smart city environments.