WOS+: Wait-for-Optimal-Set for Wireless Federated Learning
Abstract
The growing need for privacy-preserving machine learning has catalyzed research into federated learning (FL). Established FL client-selection schemes exhibit distinct limitations: “Wait-for-all” suffers from straggler effects due to heterogeneous client computation times, while “Wait-for-one” converges slowly because clients contribute highly diverse and outdated model versions. “Wait-for-<inline-formula> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula>” combines the two schemes by allowing the central server to select <inline-formula> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula> clients in each round, while the remaining clients continue local training. However, this approach may not work well in practice since determining an appropriate value of <inline-formula> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula>—whether fixed or dynamically adjusted—is fundamentally difficult due to the tight coupling between the effective round duration, wireless channel conditions, and scheduling policy. Therefore, instead of fixating on finding a value <inline-formula> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula>, this paper proposes a novel client selection and aggregation strategy called Wait-for-Optimal-Set (WOS+) that dynamically determines the cut-off time that selects an optimal subset of clients to participate in each round. WOS+ further integrates a joint resource-block (RB) allocation, modulation-and-coding-scheme (MCS) selection, and power-allocation algorithm to enable immediate uplink transmission upon completing local training at a client. Through extensive simulations in a UAV-based FL scenario, WOS+ demonstrates superior convergence speed and model accuracy compared to existing methods, especially under time-varying channels and heterogeneous device capabilities.