The rapid growth and adoption of the Internet of Things (IoT) have led to ecosystems comprising interconnected devices that can generate and exchange vast amounts of data. However, as promising as the intelligent applications enabled by these devices are, they pose significant challenges, particularly in terms of limited computational resources and the need for timely and accurate decision-making. In this context, we propose a Genetic Algorithm (GA)-based optimization algorithm for hyperparameter tuning of Machine Learning (ML) models aligned with the resource constraints and rapid decision-making demands of IoT. The introduced Time-Sensitive Genetic Algorithm (TSGA) jointly optimizes both the predictive accuracy and the execution time of the ML model. Furthermore, we introduce a novel Data Thinning (DT) component, which achieves a considerable reduction in the input data volume without compromising the model’s accuracy. The component is built upon the Largest-Triangle-Three- Buckets LTTB algorithm and an adaptive oine cumulative sum (CUSUM)-based test. We validated TSGA and DT on real IoT data traces, and the results demonstrate that TSGA can determine hyperparameter settings that strategically balance the accuracy metrics of interest with execution time while reducing the data volume burden. Our approach generally outperforms baseline methods in both predictive performance and runtime responsiveness, with only a few exceptions where a slight decrease in accuracy yields substantial gains in execution time, proving its value for resource-constrained, real-world applications.
Loading....