Enhanced Optimal Stopping Algorithm-Assisted Node Selection in IoT Network for Federated Learning

Abstract

Federated Learning in Internet of Things (IoT) networks involves a critical trade-off between resource efficiency and global model performance. We propose a lightweight optimal stopping theory-assisted node selection strategy, which unifies local model accuracy (LMA) and received signal strength (RSS) into a single scoring metric governed by a tunable parameter α. The optimized value α maximizing the probability of picking the best elements following our strategy has been obtained through a probabilistic framework. The analytical results have been confirmed through simulation. Experiments have been conducted employing the MNIST dataset and show consistent improvements in global model accuracy compared to four baseline strategies, while ensuring a suitable complexity for real-time IoT deployment. Moreover, unlike clustering-based or secretary problem-inspired approaches, our method evaluates all nodes globally and performs reliably even with limited candidate availability.