Leveraging Fully-decoupled Radio Access Network for Wireless Federated Learning Acceleration
Yunting Xu, Luofang Jiao, Tianqi Zhang, Haibo Zhou, Sherman Shen · 2022 IEEE/CIC International Conference on Communications in China (ICCC) · 2022
Federated learning (FL) has emerged as an innovative machine learning (ML) paradigm that can utilize the computation capability of both the cloud and the end-users to support data-intensive tasks in the next-generation mobile communication networks. However, the collaboration between the cloud and the end-users is usually constrained by the worst wireless link quality during the uplink and downlink transmission. In this paper, targeting at reducing the FL training latency over the wireless networks, we leverage the uplink and downlink fully-decoupled radio access network (FD-RAN) architecture through the coordinated multiple base stations (BSs) access mechanism. First, based on a proven data rate lower bound, we exploit the integer variable relaxation and the successive convex approximation (SCA) algorithms to transform the original computationally prohibitive non-convex problem into a solvable convex form. Subsequently, the Lagrange dual decomposition method is used to obtain an optimal BS serving cluster (OSC) for the uplink and downlink transmission respectively. Extensive simulations are conducted to verify the effectiveness of the proposed coordinated multiple BSs access solution for realizing a faster FL training task.