Training Efficiency of Federated Learning: A Wireless Communication Perspective
Shunan Yang, Yuan Liu · 2020
In this paper, we optimize federated learning at network edge via radio resource allocation and scheduling, including user selection, bandwidth allocation and batch-size allocation. We define a new performance metric, namely training efficiency, to not only accelerate the convergence but also increase the accuracy of the training process. The studied problem is non-convex and we develop an efficient algorithm to solve it. The proposed policy adapts wireless channels, computing capacities, and local datasets of the users. The system performance improvement contributed by the proposed scheme is demonstrated by experiments.