classifier gets a small weight. With this in mind, some authors have developed their own
heuristics for weighting the ensemble members. For binary classification problems, Freund
et al [39] weight predictors exponentially with respect to their training error. Tresp and
Taniguchi [134] investigate non-constant weighting functions, where the weightings are determined anew for each testing pattern, by calculating a probability for how likely a given
estimator is to have seen data in a region of the input space close to a new input pattern.