Automated Support Vector Machine Kernel Function Selection Based on Bayesian Optimization for Prognostic Learning Path Loss Datasets

Abstract

 Precise conceptualisation of signal loss in environments with varying terrains, urban settings, and other factors are essential for network planning and optimization. Support Vector Machines (SVMs) have emerged as powerful tools for path loss prediction due to their strong generalization capabilities. However, the performance of an SVM is highly dependent on the judicious selection of its kernel function and corresponding hyperparameters. Traditional methods for kernel selection, such as manual trial-and-error or exhaustive grid search, are computationally expensive, time-consuming, and often sub-optimal, especially given the complex and non-linear nature of path loss phenomena. This paper proposes a novel automated framework for simultaneous SVM kernel function selection and hyperparameter optimization using Bayesian Optimisation (BO). Leveraging BO's efficiency in exploring complex, high-dimensional search spaces with expensive objective functions, our method systematically identifies the optimal kernel type and its associated parameters, tailored for specific path loss datasets. The optimized SVMs are then integrated into a prognostic learning framework to forecast future channel conditions, anticipate link degradation, and enable proactive resource management. Experimental evaluation across diverse field path loss datasets demonstrates that the proposed BO-driven approach significantly improves prediction accuracy and computational efficiency, paving the way for more robust dynamic wireless communication systems