Performance evaluation of machine learning methods for path loss prediction in rural environment at 3.7 GHz

This paper presents and assesses various machine learning methods that aim at predicting path loss in rural environment. For this purpose, models such as artificial neural network (ANN), support vector regression (SVR), random forest (RF), and bagging with k-nearest neighbor (B-kNN) learners, are exploited and evaluated. They are trained and tested with path loss data collected from an extensive measurement campaign that have been carried out in diverse rural areas in Greece. The results demonstrate that all the proposed machine learning models outperform the empirical ones, exhibiting, in any case, root-mean-square-error (RMSE) values between 4.0 and 6.5 dB. The poorest prediction of the measured data is encountered for SVR with Polynomial kernel. Furthermore, B-kNN and RF algorithms preserve comparable path loss approximations with remarkably low RMSE on the order of 4.2–4.3 dB. The error metrics also reveal that increasing the number of hidden layers in ANNs, their performance is gradually enhanced. However, deeper layouts with more than three hidden layers do not markedly improve any further the prediction accuracy. Finally, the best prediction is achieved when employing a three-hidden layered ANN with 51 neurons evenly distributed among the layers. The specific layout exhibits the lowest RMSE value (4.0 dB), thus being highly recommended for accurate path loss predictions in rural locations.