CARAC-MDRL: Cluster-based Ant-colony Routing with Adaptive Cluster Head and Multi-Agent Deep Reinforcement Learning

VANETs face major challenges related to rapid mobility, topological variability, and quality of service requirements. To address these issues, this paper proposes CARAC-MDRL, an intelligent extension of the CARAC protocol, aiming to enhance cluster stability and routing efficiency. CARAC-MDRL introduces a dual Deep Reinforcement Learning (DRL) agent architecture: a CH-DRL agent deployed at the Cluster Heads to dynamically regulate intra-cluster stability parameters, and RSU-DRL agent located at the Road Side Units to globally adjust the ant colony optimization parameters. These agents cooperate via a hierarchical federated learning mechanism, enabling collective network adaptation without exchanging raw data. Simulations conducted using NS-3 demonstrate that CARAC-MDRL outperforms CARAC, AQRV, and CPB in terms of latency, throughput, and packet delivery ratio. By combining the structural robustness of CARAC with the adaptive intelligence of reinforcement learning, this approach offers a scalable and high-performance solution for routing in dynamic urban VANET environments.