Optimal Stopping-Based Online Node Selection for Multi-Parameter Federated Learning in IoT Networks Under Multiuser Interference

Federated Learning (FL) is a promising approach that enables multiple devices to collaboratively train a machine learning model over a wireless IoT network. It maintains data privacy by providing only model updates, rather than the original data. Nevertheless, the primary challenge remains in the selection of the most suitable clients to participate in each training round, which is influenced by resource constraints and the quality of the central entity–client link, including multiuser interference (MUI). Existing FL node selection schemes either rely on single-metric heuristics (e.g., based solely on local model accuracy or on link quality) or on complex Reinforcement learning (RL)/AutoFL controllers that are difficult to train and analyze under wireless interference. This leaves a gap for lightweight, analytically tractable schemes that jointly exploit model quality and link reliability in interfered IoT networks. In the presence of MUI at each device, this paper exploits optimal stopping theory (OST) to develop a novel online node selection strategy that combines the local model accuracy (LMA) and the received Signal-to-Interference-plus-Noise Ratio (SINR) into a single metric with an adjustable weight α. The system mandates that the LMA and SINR of each selected node exceed specific thresholds. The probability of selecting at leastmnodes from the existing set is maximized by conducting a comprehensive analysis to determine the optimal parameter α, whereas an extensive discussion regarding the impact of various parameters on the system’s efficiency has been provided. The proposed OST-based approach obtains superior global model accuracy (GMA) and convergence in comparison to baseline methods, with a lower system complexity, as evidenced by semi-synthetic experiments that utilize the CRAWDAD kth/rss dataset.