Stochastic Device Scheduling and Power Control in Federated Learning with Energy Harvesting

Federated Learning (FL) is emerging as a promising approach for enabling collaborative machine learning without centralizing data, particularly within Internet of Things (IoT). However, optimizing FL efficiently in dynamic environments under delay and energy constraints is challenging due to factors such as device heterogeneity and various connection quality. Moreover, the energy harvesting capabilities of many IoT devices add a new dimension to optimizing FL in wireless networks, a topic that remains underexplored. In this paper, we formulate an optimization problem to balance the trade-off between maximizing device scheduling probability and minimizing communication overhead. By leveraging Lyapunov optimization, we transform the original objective into a stability problem. We then introduce Sto-POLISH, a stochastic power control and device scheduling algorithm, that dynamically selects optimal power and scheduling probabilities while satisfying three long-term constraints. Simulations demonstrate that Sto-POLISH shows great performance in achieving fast convergence, reducing battery energy consumption, and maintaining communication overhead within acceptable limits compared to existing benchmarks.