AMAGNN: A High-Performance Multi-Agent GNN Framework for Scalable Congestion Control in VANETs

Traditional congestion control mechanisms in VANETs are affected by dynamic topologies and traffic fluctuations to generate critical challenges. These methods often fail to generalize under high-mobility conditions and heterogeneous traffic densities. To address these limitations, we propose the Adaptive Multi-Agent Graph Neural Network (AMAGNN), a novel framework that integrates Graph Neural Networks (GNNs) with Multi-Agent Reinforcement Learning (MARL) to enable distributed and cooperative congestion control in VANETs. AMAGNN models the vehicular environment as a temporally evolving graph, where each vehicle operates as a decentralized intelligent agent. These agents leverage local and neighbourhood state information, encoded via GNNs, to learn optimal routing and communication strategies through MARL.Additionally, we evaluate the performance of a Swarm Intelligence-Based Distributed Congestion Control (Swarm-VANET) approach that employs bio-inspired optimization techniques—specifically Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO)—to adaptively manage routing and transmission parameters based on local traffic conditions.Empirical results demonstrate that AMAGNN significantly outperforms conventional baselines such as GCN-RL, DQN-Adaptive, MA-PPO, and Swarm-VANET. AMAGNN achieves a Packet Delivery Ratio (PDR) of 92.5%, average end-to-end delay of 28.4 ms, and throughput of 275.6 kbps, underscoring its robustness and reliability. Furthermore, it achieves the lowest congestion index (0.35) and average waiting time (0.31 s), validating its effectiveness in mitigating traffic bottlenecks. These results highlight the superiority of AMAGNN in enhancing data dissemination efficiency, scalability, and real-time adaptability in dynamic VANET environments.