A class of regression models taking into account the lack of sufficient statistical data is proposed. The main ideas of the class are to use the framework of Vapnik's learning theory and to replace a single probability distribution of the noise by a set of probability distributions, which is not...