A Greedy Control Policy for Latency and Energy Constrained Wireless Federated Learning

Abstract

For federated learning, each device participates in the model learning in a collaborative training manner. Due to the constraint of delay and energy consumption in actual wireless environment, resource allocation is essential for the convergence speed of federated learning. This paper analyzes the convergence bound of federated learning from a theoretical perspective, based on which, we propose a greedy control policy that combines aggregation frequency control and device scheduling together. The proposed policy minimizes the loss of model training under a given time and energy budget with a greedy strategy which eliminates the device with the worst performance gain in each step. Simulation results show that under different wireless environments, the proposed global control policy achieves higher accuracy than the commonly used federated learning algorithms and has a good robustness to non-i.i.d. data.