Scholarlink

Purpose <p>Congestion control in Vehicular Ad-hoc Networks (VANETs) remains challenging due to highly dynamic topologies, heterogeneous communication modes, and fluctuating vehicular densities. Conventional heuristic and reinforcement learning approaches often fail to achieve scalable, adaptive, and cooperative decision-making under such dynamic conditions.</p> Methods <p>This study proposes SAGE-Fed TL-MARL — a <i>Semantic-Aware</i>,<i> Graph-based</i>,<i> Event-triggered Federated Transfer Learning Multi-Agent Reinforcement Learning</i> framework for adaptive congestion control in VANETs. Each Roadside Unit (RSU) extracts high-level semantic embeddings (congestion likelihood, queue state, and uncertainty) using lightweight deep encoders, which are integrated into a dynamic vehicular graph. Graph Neural Network (GNN)-based MARL agents collaboratively optimize routing, transmission power, and scheduling decisions. An event-triggered federated communication mechanism enables selective prototype sharing among RSUs, ensuring scalability, privacy preservation, and reduced synchronization overhead.</p> Result <p>Extensive SUMO–OMNeT + + co-simulations across three urban networks—Singapore CBD, Chennai Inner Ring Road, and Hyderabad Outer Corridor—show that SAGE-Fed TL-MARL consistently outperforms state-of-the-art baselines, including AODV, GNN-MARL, TL-MARL, Non-transfer MARL, SARSA-RL, and DCRACC. The proposed framework achieves up to 18–25% higher Packet Delivery Ratio (PDR), 20–30% lower delay, 15–25% lower congestion index, and 35–45% reduced routing overhead, while maintaining a fairness index above 0.9 across all topologies.</p> Conclusion <p>The findings confirm that SAGE-Fed TL-MARL provides a scalable, privacy-preserving, and semantically intelligent solution for congestion control in next-generation VANETs. Its strong performance across heterogeneous city environments demonstrates superior generalization, adaptability, and communication efficiency, representing a significant advancement over existing state-of-the-art methods.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Semantic-aware, graph-based, event-triggered federated transfer-learning multi-agent reinforcement learning (SAGE-Fed TL-MARL) for congestion control in VANETs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月20日

查看全文
  • Prashanta Kumar Patra,
  • Santosh Kumar Maharana

摘要

Purpose

Congestion control in Vehicular Ad-hoc Networks (VANETs) remains challenging due to highly dynamic topologies, heterogeneous communication modes, and fluctuating vehicular densities. Conventional heuristic and reinforcement learning approaches often fail to achieve scalable, adaptive, and cooperative decision-making under such dynamic conditions.

Methods

This study proposes SAGE-Fed TL-MARL — a Semantic-Aware, Graph-based, Event-triggered Federated Transfer Learning Multi-Agent Reinforcement Learning framework for adaptive congestion control in VANETs. Each Roadside Unit (RSU) extracts high-level semantic embeddings (congestion likelihood, queue state, and uncertainty) using lightweight deep encoders, which are integrated into a dynamic vehicular graph. Graph Neural Network (GNN)-based MARL agents collaboratively optimize routing, transmission power, and scheduling decisions. An event-triggered federated communication mechanism enables selective prototype sharing among RSUs, ensuring scalability, privacy preservation, and reduced synchronization overhead.

Result

Extensive SUMO–OMNeT + + co-simulations across three urban networks—Singapore CBD, Chennai Inner Ring Road, and Hyderabad Outer Corridor—show that SAGE-Fed TL-MARL consistently outperforms state-of-the-art baselines, including AODV, GNN-MARL, TL-MARL, Non-transfer MARL, SARSA-RL, and DCRACC. The proposed framework achieves up to 18–25% higher Packet Delivery Ratio (PDR), 20–30% lower delay, 15–25% lower congestion index, and 35–45% reduced routing overhead, while maintaining a fairness index above 0.9 across all topologies.

Conclusion

The findings confirm that SAGE-Fed TL-MARL provides a scalable, privacy-preserving, and semantically intelligent solution for congestion control in next-generation VANETs. Its strong performance across heterogeneous city environments demonstrates superior generalization, adaptability, and communication efficiency, representing a significant advancement over existing state-of-the-art methods.

Graphical abstract