Enabling Efficient Scheduling Policy in Intelligent Reflecting Surface Aided Federated Learning

Abstract

Federated learning (FL) has been proposed to coordinate multiple edge user equipments (UEs) for training a global model. However, the FL's performance is affected by the channel state of wireless network. Specifically, the performance of the random selecting UE is seriously limited by the millimeter-wave (mmWave) channel. In this paper, we propose to deploy intelligent reflecting surface (IRS) to reconstruct the non-line-of-sight (NLoS) mmWave channel. A novel UE scheduling strategy is then proposed to optimize the FL system performance. For selecting the particular UEs and achieving higher convergence, we formulate an optimization problem that jointly optimizes the aggregation vector of the base station (BS) and the IRS phase shift matrix. To figure out the formulated problem, we propose a two-step difference-of-convex (DC) algorithm. Simulation results demonstrate that the proposed algorithm can achieve higher convergence and a lower training loss than the benchmark algorithm.