Adaptive Congestion Control in VANET Using MARL Based Transfer Learning Technique

Vehicular Ad-hoc Networks (VANETs) play a vital role in Intelligent Transportation Systems (ITS) by enabling real-time communication between vehicles and infrastructure. However, high mobility, dynamic topology, intermittent connectivity, and limited bandwidth make congestion control a complex task. Traditional methods like static routing and rule-based control lack adaptability in such environments. Recent advances in deep reinforcement learning (DRL) and multi-agent reinforcement learning (MARL) offer promising alternatives by enabling vehicles to learn adaptive strategies through interaction between the nodes (vehicles, Road Side Units (RSUs), On-Board-Units (OBUs)). However, these approaches face challenges including slow convergence due to sparse and delayed rewards, poor scalability in large networks due to increased agent coordination overhead, and limited generalization, as policies trained in one scenario often fail in new conditions. To address these issues, we propose a transfer learning (TL)-based MARL framework for adaptive congestion control in VANETs. The proposed method transfers knowledge from pretrained policies and fine-tunes them in new target scenarios, improving learning efficiency, scalability, and adaptability. This reuse of prior knowledge allows faster convergence and reduces training overhead while maintaining high performance under varying conditions. Experimental results demonstrate that the proposed TL-MARL framework outperforms existing approaches such as conventional MARL, Federated Graph Neural Network-based Multi-Agent Reinforcement Learning (FGNN-MARL), and Graph Convolutional Network-based Multi-Objective Multi-Agent Reinforcement Learning (GCN-MO-MARL). The proposed method achieves a 92.5% packet delivery ratio, 28.4 ms average end-to-end delay, and 275.6 kbps throughput, while reducing packet loss by over 50% and decreasing the congestion index by 32% compared to non-transfer MARL models. Ablation and sensitivity analyses confirm the robustness of framework and its ability to generalize across diverse network topologies and traffic conditions.