Machine Learning models for the prediction of Wi-Fi links performance using a CityLab testbed

The Wi-Fi links performance depends in a highly complex way on the actual topology, channel qualities, spectral configurations, etc. Existing Wi-Fi radio link performance models usually adopt explicit and bottom-up approaches in order to predict throughput figures bawd on Markov chains and SINR levels. In this work we have validated a new approach for predicting the performance of Wi-Fi networks. Based on data measurements from the outdoor Wi-Fi CityLab testbed in Antwerp we have tested four different supervised learning algorithms. We observed that abstract "black box" models built using supervised machine learning techniques — without any deep knowledge of the complex interference dynamics of IEEE 802.11 networks — can estimate the link throughput with very good accuracy, reaching a value of R2-score of 90% for the case of the Gradient Boosting Regressor.