Mobility-Aware Federated Learning: Multi-Armed Bandit Based Selection in Vehicular Network

Data-driven machine learning (ML) tasks in vehicular networks In this paper, we design a Mobility-Aware Vehicular such as trajectory prediction, object detection and traffic Federated Learning (MAVFL) scheme, where vehicles drive sign classification enhance road safety and alleviate urban through the road segment to participate in FL with a collaborative congestion to facilitate autonomous driving [1]. The distributed base station (BS). We propose the real-time ratio data of each vehicle is collected by various sensors such as that vehicles successfully upload models. We conduct the GPS (Global Positioning System), LiDAR (Light Detection theoretical analysis of convergence and demonstrate that the and Ranging) and cameras, and increased data privacy and ratio significantly influences convergence. Based on analytical communication overhead is brought in when local data is offloaded results, we formulate the optimization problem to maximize to the server. Federated learning (FL) enables vehicles the utility function while minimizing training loss and training to collaboratively train models from the server aggregated from delay. We design an MAB-based vehicle selection algorithm all vehicles without sharing local data directly and reduces to solve the optimization problem. Extensive simulation results the communication overhead caused by large amounts of data show the effectiveness of the proposed scheme in terms of transmission between vehicles to the server and cloud [2], [3].