No label defined (Q3159491)

This work has three parts: theoretical foundations of Monte-Carlo methods for self-learning of Gaussian mixed distributions, development of these methods for augmentation and reduction of Gaussian mixture models, and validation of all obtained results. The first part analyzes factors responsible for self-learning of Gaussian mixed distributions and shows how to control these factors in order to generalize effectively. This part describes self-learning as a balance problem between the number of training data and the probability measure that should be minimized. The second part considers a new statement of the self-learning problem of Gaussian mixed distributions. The problem of estimating parameters of a function at given points of interest is introduced. Therefore, new algorithms for direct estimation of the parameters of Gaussian mixed distributions are given. Using these methods, a modification of learning data is obtained. In the last part, all obtained results are compared with expectation-maximization algorithms. Convergence of these results forms the basics for the theory of self-learning processes of Gaussian mixed distributions. Bibliographical, general comments reflect the author's point of view on the development of these topics.