Accurate estimation of wireless propagation characteristics is essential for guiding the design and deployment of fifth-generation (5G) communication systems. As network demand increases and 5G infrastructure is introduced in progressive phases, reliable path loss (PL) prediction models are required to refine deployment strategies and improve network efficiency. Conventional propagation models frequently display limited flexibility when applied to diverse environmental conditions and often entail considerable computational expense, reducing their practicality for large-scale 5G planning. Recent developments in data-centric artificial intelligence (AI) have enabled more adaptive and analytically powerful approaches to propagation modeling, resulting in notable gains in PL prediction accuracyThis study employs a comprehensive dataset produced using the NYUSIM channel simulator, integrating a wide spectrum of atmospheric parameters and seasonal variations within South Asian urban microcell environments, complemented by broad empirical observations. The core objective is to construct, optimize, and evaluate four machine learning (ML) models capable of accurately predicting PL at high-frequency bands critical to 5G performance. A fully automated hyperparameter tuning pipeline, based on the Optuna framework, is applied to twelve regression algorithms, including advanced ensemble methods, regularized linear techniques, and classical baseline models. Performance assessment emphasizes predictive reliability, stability, and cross-model generalization. Furthermore, statistical analysis utilizing bootstrap confidence intervals and paired t-tests indicates that all ML methods perform equivalently (p > 0.4), while SHapley Additive exPlanations (SHAP) analysis across all models supports a consistent feature importance distribution, supporting the statistical analysis results. To showcase the superiority of the ML approaches, a comparison with conventional free-space PL modeling methods is presented, with the AI methodology demonstrating robust performance across seasonal variations and a 95.3% improvement.
Rekkas et al. (Mon,) studied this question.