Ensembling Probabilistic Regressors for Path Loss Prediction

Machine Learning (ML) based Path Loss (PL) prediction is of great importance when designing modern communications networks. As a result of providing a framework for understanding what learning entails, probabilistic modeling has become one of the most prominent theoretical and practical methods for developing machines that acquire knowledge from experience-based data. Our work focuses on combining two probabilistic regressors, namely NGBoost and XGBoost-Distribution, into a stacked generalization ensemble. As opposed to the vast majority of stacked ensemble regressors, the novelty of our proposed ensemble is that each base learner delivers two, instead of one, outputs to the meta-learner. That is, apart from the predicted path loss value, each probabilistic base learner computes the standard deviation of the corresponding prediction. Results show that these additional inputs of the meta-learner (which are the outputs of the base learners) pave the way for more accurate predictions.