Quantifying Prediction Uncertainty in Regression using Random Fuzzy Sets: the ENNreg model

We introduce a neural network model for regression in which prediction uncertainty is quantified by Gaussian random fuzzy numbers (GRFNs), newly introduced family of subsets the real line that generalizes both variables and possibility distributions. The output GRFN constructed combining GRFNs induced prototypes using combination operator Dempster's rule Evidence Theory. three units indicate most plausible value response variable, variability around this value, epistemic uncertainty. trained minimizing loss function negative log-likelihood. Comparative experiments show method competitive, terms accuracy calibration error, with state-of-the-art techniques such as forests or deep learning Monte Carlo dropout. In addition, outputs predictive belief can be shown to calibrated, sense it allows us compute conservative intervals specified degree.