Hybrid NOMA for Time-Sensitive Wireless Federated Learning Through Lyapunov Optimization Based Resource Scheduling

Federated Learning (FL) has gained prominence as a decentralized approach for training a shared machine learning model across multiple participants without compromising their private raw data. Nevertheless, integrating such a framework into wireless network presents unique challenges, including limited radio resources and heterogeneous participant resource configurations. To tackle these issues, this paper explores the potential of Non-Orthogonal Multiple Access (NOMA), enabling simultaneous user access to a shared Resource Block (RB). Considering the introduced delay due to Successive Interference Cancellation (SIC) in pure NOMA, an uplink hybrid NOMA scheme is introduced in this study. The goal is to reduce the overall task time within the energy constraint of each participant, incorporating dynamic user selection and resource allocation strategies while ensuring FL convergence. This is facilitated through a Lyapunov optimization technique that simplifies the complex, long-term optimization into a tractable, round-by-round online decision-making algorithm. Then a novel matching theory-based method is introduced for joint user selection and pairing, which is accompanied by a local resource allocation algorithm. Simulation results validate the efficacy of our proposed approach, demonstrating the reduction in FL training duration without compromising on convergence integrity under diverse scenarios.