Development of Machine Learning-Based Radio Propagation Models and Benchmarking for Mobile Networks

Abstract

Path loss prediction plays an important role in assisting network operators to build quality mobile networks by indicating optimal coverage and interference performances. Common empirical models such as the Cost 231-Hata model and Cost 231-Walfisch Ikegami model have been used to predict path loss due to their simplicity and robustness; however, these models do not provide reliable accuracy for small coverage infrastructures, especially in LTE and 5G networks. Therefore, as an alternative method, deterministic models (e.g., ray-tracing models) are widely used for small coverage applications–however, these also require precise 3D digital maps, street structures, and building surface reflection indices, etc. which can be costly and need frequent maintenance. In this paper, an alternative model based on machine learning algorithms is proposed and the algorithms' prediction accuracies are evaluated. The machine learning approach trains models using signal strength data upon which they draw generalizations. The performance of these models is compared to that of the empirical models, and the results reveal that the machine learning models perform significantly better, with Random Forest outperforming other algorithms.