AI Boosts Low-Latency Relay Selection for V2X Communications

Reliable and low-latency uplink connectivity is crucial for C-V2X networks, especially in dense urban areas where direct vehicle-to-infrastructure links are often degraded. Multi-hop relaying can restore coverage, but selecting optimal relays under various constraints is an NP-hard problem, typically solved by computationally intensive Mixed-Integer Linear Programming (MILP). This paper introduces a novel approach to address this challenge. The proposed framework, called edge-aware Learning-to-Optimise, uses Graph Isomorphism Networks with Edge Features (GINE) for real-time relay selection. It models each V2X scenario as a directed graph, incorporating vehicle state, traffic demand, and radio-link capacity as node and edge features. An offline MILP oracle trains the GINE model, allowing it to predict optimal relay configurations through a single, low-latency forward pass. Experiments using a large-scale dataset demonstrated that GINE closely matches MILP decisions, achieving high accuracy and F1-scores. It also provides significant end-to-end connectivity gains over baseline methods, with inference latency consistently below 5 milliseconds. A hybrid GINE-Pruned MILP (GP-MILP) strategy further ensures MILP-equivalent solutions while keeping solver runtimes within stringent NR-V2X latency budgets, making advanced optimization practical for vehicular communications.