Censored Path Loss Prediction With Stacked Ensemble of Deep Neural Networks

This paper proposes a stacked ensemble of deep neural networks designed to predict true path loss values that are unobservable beyond the censoring threshold, which is determined by the receiver sensitivity. The proposed model consists of five Censored Regression Neural Networks (CRNNs) and a Blending Neural Network (BNN). Each CRNN is trained on censored path loss samples by minimizing a likelihood-based loss function. Although trained with censored samples, the designed loss function enables the prediction of true path loss values beyond the threshold that limits the path loss observations. The BNN provides the final prediction by optimally integrating the outputs of the CRNNs. This prediction mechanism facilitates accurate estimation beyond the threshold by mitigating prediction bias, whereas conventional machine learning-based path loss models yield biased predictions below the threshold.