Maximum downlink throughput prediction in 5G networks

Abstract

In this work, an experiment to examine the feasibility of maximum downlink throughput prediction in 5G networks through machine learning is conducted. The goal is to find a generalisable methodology capable of predicting the throughput at any given time and location. The approach could be further utilised to aid in searching for throughput testing locations of mobile devices in commercial 5G networks.

These studies are often restricted by the lack of diversity of the datasets due to the time-consuming and financially demanding nature of comprehensive data collection. The experiment presented in this thesis utilises a diverse dataset collected with Mediatek's in-house tool Field Capture System and various machine learning models to find the optimal algorithm for the given task. The performance of the models is evaluated with metrics such as coefficient of determination, mean squared error and mean absolute error across various subsets of data to gain further understanding of the suitability of the distinct models.

The best models, AdaBoost and random forest regressor, achieved performance that is both in line with previous related studies and also able to highlight the suitability of such regression algorithms in this unique setting. Achieved \(R^2\) scores of 0.89 and 0.88, respectively, prove that sufficient performance was obtained in this complex task, despite the few but notable limitations of the used dataset. The methods and findings of this thesis successfully lay a further baseline for future research and applications.