Straggler Effect Mitigation for Federated Learning in Cell-Free Massive MIMO

Thanh Tung Vu, Duy T. Ngo, Hien-Quoc Ngo, Dao N. Minh, Tran H. Nguyen, Middleton H. Richard

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Straggler effect is the main bottleneck in realizing federated learning (FL) in wireless networks. This work proposes a novel user (UE) selection approach to mitigate this effect with UE sampling in cell-free massive multiple-input multiple-output networks. Our proposed approach selects only a small subset of UEs for participating in one FL process. Importantly, since the UEs are selected before any FL process is executed, the performance of FL during the executing time is not affected by our method. Here, we select UEs by solving an FL transmission time
minimization problem that jointly optimizes UE selection, power control, and data rate. The problem is formulated to capture the complex interactions among the FL training time, UE selection, and straggler effect. This mixed-integer mixed-timescale stochastic nonconvex problem is constrained by the minimum number of UEs to guarantee the quality of learning. By employing online successive convex approximation, we propose a novel algorithm to solve the formulated problem with guaranteed convergence to the neighbourhood of their stationary points. Our approach can significantly reduce the FL transmission time over baseline
approaches, especially in the networks that experience serious straggler effect due to the moderately low density of access points.
Original languageEnglish
Title of host publicationIEEE International Conference on Communications
Publisher IEEE
Publication statusAccepted - Jan 2021

Fingerprint

Dive into the research topics of 'Straggler Effect Mitigation for Federated Learning in Cell-Free Massive MIMO'. Together they form a unique fingerprint.

Cite this