Federated Learning in Multi-antenna Wireless Networks

In this work, we propose an analytical model to study the convergence performance of federated learning (FL) in a multi-antenna wireless network by comparing three different scheduling policies, i.e., the round robin (RR), random scheduling (RS), and proportional fair (PF). We derive tractable expressions for the convergence rates of FL by taking into account the scheduling policy, antenna number, channel fading, and inter-cell interference. Our results show that PF achieves the best convergence performance under the high signal-to-interference-plus-noise ratio (SINR) threshold, while the three scheduling policies achieve very similar convergence rates under extremely low SINR threshold. Given the number of antennas per base station (BS), we observe an optimal number of scheduled user equipments (UEs) per BS that maximizes the convergence rate of FL under high SINR threshold because of the trade-off between serving more UEs and achieving higher successful transmission probability.