A Novel Hybrid NOMA-TDMA Scheme for Wireless Federated Learning Networks

Abstract

To address the straggler problem in wireless federated learning (WFL) networks, a novel hybrid non-orthogonal multiple access (NOMA) - time division multiple access (TDMA) scheme is proposed, where not only local model training and uploading of different users are performed in parallel, but also the model uploading of different users is performed in a hybrid NOMA-TDMA manner. Particularly, users that complete their local training form dynamic NOMA groups across different time slots in a TDMA fashion to upload their models efficiently. The latency per training round minimization problem is formulated to jointly optimize the user scheduling, time allocation, computing frequency allocation, and energy allocation. An efficient algorithm, combining one-dimensional search and successive convex approximation, is developed. Simulation results demonstrate that the proposed scheme can outperform the state-of-the-art NOMA and TDMA schemes under various system parameter setups.