TY - JOUR
T1 - Multi-armed bandit-based client scheduling for federated learning
AU - Xia, Wenchao
AU - Quek, Tony Q.S.
AU - Guo, Kun
AU - Wen, Wanli
AU - Yang, Howard H.
AU - Zhu, Hongbo
N1 - Funding Information:
Manuscript received December 8, 2019; revised April 1, 2020 and June 17, 2020; accepted July 5, 2020. Date of publication July 16, 2020; date of current version November 11, 2020. This work was supported in part by the Singapore University of Technology and Design (SUTD) Growth Plan Grant for AI, in part by the SUTD-ZJU Seed under Grant 201909, and in part by the National Natural Science Foundation of China under Grant 61871446. The associate editor coordinating the review of this article and approving it for publication was S. Dey. (Corresponding author: Hongbo Zhu.) Wenchao Xia, Tony Q. S. Quek, Kun Guo, Wanli Wen, and Howard H. Yang are with the Information Systems Technology and Design Pillar, Singapore University of Technology and Design, Singapore 487372 (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]).
Publisher Copyright:
© 2002-2012 IEEE.
PY - 2020/11
Y1 - 2020/11
N2 - By exploiting the computing power and local data of distributed clients, federated learning (FL) features ubiquitous properties such as reduction of communication overhead and preserving data privacy. In each communication round of FL, the clients update local models based on their own data and upload their local updates via wireless channels. However, latency caused by hundreds to thousands of communication rounds remains a bottleneck in FL. To minimize the training latency, this work provides a multi-armed bandit-based framework for online client scheduling (CS) in FL without knowing wireless channel state information and statistical characteristics of clients. Firstly, we propose a CS algorithm based on the upper confidence bound policy (CS-UCB) for ideal scenarios where local datasets of clients are independent and identically distributed (i.i.d.) and balanced. An upper bound of the expected performance regret of the proposed CS-UCB algorithm is provided, which indicates that the regret grows logarithmically over communication rounds. Then, to address non-ideal scenarios with non-i.i.d. and unbalanced properties of local datasets and varying availability of clients, we further propose a CS algorithm based on the UCB policy and virtual queue technique (CS-UCB-Q). An upper bound is also derived, which shows that the expected performance regret of the proposed CS-UCB-Q algorithm can have a sub-linear growth over communication rounds under certain conditions. Besides, the convergence performance of FL training is also analyzed. Finally, simulation results validate the efficiency of the proposed algorithms.
AB - By exploiting the computing power and local data of distributed clients, federated learning (FL) features ubiquitous properties such as reduction of communication overhead and preserving data privacy. In each communication round of FL, the clients update local models based on their own data and upload their local updates via wireless channels. However, latency caused by hundreds to thousands of communication rounds remains a bottleneck in FL. To minimize the training latency, this work provides a multi-armed bandit-based framework for online client scheduling (CS) in FL without knowing wireless channel state information and statistical characteristics of clients. Firstly, we propose a CS algorithm based on the upper confidence bound policy (CS-UCB) for ideal scenarios where local datasets of clients are independent and identically distributed (i.i.d.) and balanced. An upper bound of the expected performance regret of the proposed CS-UCB algorithm is provided, which indicates that the regret grows logarithmically over communication rounds. Then, to address non-ideal scenarios with non-i.i.d. and unbalanced properties of local datasets and varying availability of clients, we further propose a CS algorithm based on the UCB policy and virtual queue technique (CS-UCB-Q). An upper bound is also derived, which shows that the expected performance regret of the proposed CS-UCB-Q algorithm can have a sub-linear growth over communication rounds under certain conditions. Besides, the convergence performance of FL training is also analyzed. Finally, simulation results validate the efficiency of the proposed algorithms.
UR - http://www.scopus.com/inward/record.url?scp=85096244689&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85096244689&partnerID=8YFLogxK
U2 - 10.1109/TWC.2020.3008091
DO - 10.1109/TWC.2020.3008091
M3 - Article
AN - SCOPUS:85096244689
SN - 1536-1276
VL - 19
SP - 7108
EP - 7123
JO - IEEE Transactions on Wireless Communications
JF - IEEE Transactions on Wireless Communications
IS - 11
M1 - 9142401
ER -