Dynamic Federated Learning Optimization: a Joint Design of Power Control and Device Scheduling

Over-the-air federated learning (OTA-FEEL) has become a communication-efficient framework for distributed machine learning that exploits the waveform superposition properties of wireless channels to aggregate gradient updates over-the-air at the parameter center, thereby improving communication efficiency. However, OTA-FEEL is inherently sensitive to wireless channel fading caused by analog transmission, which degrades learning performance. In this paper, we propose a device exploitation-exploration over-the-air (DEEOTA) FEEL method that jointly optimizes transmission power control and device scheduling to mitigate these impacts on learning performance. We apply Lyapunov framework to reformulate the derived long-term problem into multiple online subproblems based solely on instantaneous channel state information (CSI). We propose a scheduling scheme to dynamically explore and exploit numerous devices, improving model accuracy without introducing signifi-cant computational complexity.