Scholarlink

In the federated learning enabled vehicular networks, network resources are relatively limited. If all vehicles participate in model training, it will consume a large amount of network bandwidth, leading to network congestion and affecting the efficiency of data transmission. In this paper, a cluster-based federated learning approach for optimal scheduling of vehicular network user is proposed. Vehicles are clustered based on inference similarity to better learn information about the surrounding environment. From the perspective of minimizing the communication delay, the proposed method considers four key factors: data importance, communication rounds, channel quality, and single-round communication delay, to determine the optimal probability of vehicle selection. Considering that the variability of the transmission success rate of different devices can lead to the failure of timely uploading of valid local models, a weight of the success rate is set during RSU aggregation to reduce the impact of model bias. The experimental results show that the algorithm accuracy has been improved by 6% compared to traditional methods, and the delay has been reduced by 30% compared to the optimal comparison method.

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

A Vehicles Scheduling Algorithm Based on Clustering Based Federated Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xin Zhang,
  • Chi Zhang,
  • Shuyan Liu,
  • Zilong Jin

摘要

In the federated learning enabled vehicular networks, network resources are relatively limited. If all vehicles participate in model training, it will consume a large amount of network bandwidth, leading to network congestion and affecting the efficiency of data transmission. In this paper, a cluster-based federated learning approach for optimal scheduling of vehicular network user is proposed. Vehicles are clustered based on inference similarity to better learn information about the surrounding environment. From the perspective of minimizing the communication delay, the proposed method considers four key factors: data importance, communication rounds, channel quality, and single-round communication delay, to determine the optimal probability of vehicle selection. Considering that the variability of the transmission success rate of different devices can lead to the failure of timely uploading of valid local models, a weight of the success rate is set during RSU aggregation to reduce the impact of model bias. The experimental results show that the algorithm accuracy has been improved by 6% compared to traditional methods, and the delay has been reduced by 30% compared to the optimal comparison method.