Scholarlink

In vehicular edge computing (VEC), the emerging intelligent vehicles can download popular contents cached on roadside units to support real-time in-vehicle applications. As a novel distributed machine learning model, federated learning (FL) can share users’ local models to preserve privacy. However, high vehicular mobility often causes vehicles to exit VEC coverage before completing local-model uploads in conventional FL, degrading model accuracy. Asynchronous federated learning mitigates this by opportunistic aggregation, enhancing accuracy as participation grows. Furthermore, the resources of the neighboring vehicles are not considered in previous work. In this paper, we introduce a collaborative cache management (AFDQN) in vehicular network to maximize the utilization of idle resources on the road based on asynchronous federated learning and deep reinforcement learning (DRL). Vehicles train local data and upload local updates deep reinforcement learning pre-trained models to the roadside unit, which aggregates the experience of different vehicles and enhances the accuracy and adaptability of the global model. Our approach minimizes content transmission delays, thereby improving the user experience for vehicles. The method lowers average content transmission delay by 11% when compared to typical caching methods.

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

DRL-Based Asynchronous Federated Learning for Cooperative Caching in IoV

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yujie Lin,
  • Lian Tong,
  • Xiaoyu Zhu,
  • Mingfeng Su,
  • Zhiqiang Wen

摘要

In vehicular edge computing (VEC), the emerging intelligent vehicles can download popular contents cached on roadside units to support real-time in-vehicle applications. As a novel distributed machine learning model, federated learning (FL) can share users’ local models to preserve privacy. However, high vehicular mobility often causes vehicles to exit VEC coverage before completing local-model uploads in conventional FL, degrading model accuracy. Asynchronous federated learning mitigates this by opportunistic aggregation, enhancing accuracy as participation grows. Furthermore, the resources of the neighboring vehicles are not considered in previous work. In this paper, we introduce a collaborative cache management (AFDQN) in vehicular network to maximize the utilization of idle resources on the road based on asynchronous federated learning and deep reinforcement learning (DRL). Vehicles train local data and upload local updates deep reinforcement learning pre-trained models to the roadside unit, which aggregates the experience of different vehicles and enhances the accuracy and adaptability of the global model. Our approach minimizes content transmission delays, thereby improving the user experience for vehicles. The method lowers average content transmission delay by 11% when compared to typical caching methods.