Heterogeneous Training Intensity for Federated Learning: A Deep Reinforcement Learning Approach
Manying Zeng, Xiumin Wang, Weijian Pan, Pan Zhou
Abstract
Federated learning (FL) has recently received considerable attention in Internet of Things, due to its capability of letting multiple clients collaboratively train machine learning models, without sharing their private information. However, in synchronous FL, the client with weak computing or communication capability may significantly drag down the model training process, which leads to very high waiting latency for other clients. Intuitively, to alleviate this straggler problem, the clients with lower (higher) training capabilities should be assigned with less (more) training intensity. Inspired by this observation, this paper formulates a novel Heterogeneous Training Intensity assignment problem for FL, named <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HTI_FL</i> , aiming at reducing the largest training latency gap among clients. To address HTI_FL problem, we first propose an optimal deterministic algorithm, which however is only suitable for a static FL context with stable network conditions and clients' computing capabilities. To consider a practical dynamic context, we propose a Deep Reinforcement Learning Approach to learning the network conditions and clients' capabilities, and furthermore adaptively assign training intensities to clients. Finally, simulation results demonstrate the effectiveness of the proposed scheme in reducing the waiting time and accelerating the convergence of FL.