Joint Receiver Design and User Scheduling for Over-the-Air Aggregation in Federated Learning

In this article, a novel joint receiver design and user scheduling framework in over-the-air computation based federated learning (FL) system is employed to enhance the data aggregation. FL can improve the machine learning performance by updating model parameters via local training. We investigate the joint design of the user scheduling and receiver beamforming vector to minimize the distortion of the aggregated signal. To this end, we formulate a non-convex data aggregation optimization problem taking into account the impact of channel state information (CSI) and user scheduling. To solve this non-convex optimization problem, we decouple the original problem into two subproblems. First, we solve the receiver beamforming design problem for a given user scheduling result. Then, we propose a novel channel and data based user scheduling algorithm to obtain the over-the-air aggregation results. The simulation results show that the proposed scheme is more effective than the benchmark schemes.