The thesis presents the results of the evaluation of the Gaussian mixture model for the purpose of classification in operational strength. REBMIX & EM strategies for estimating the parameters of a Gaussian mixture model are presented. The parameter estimation is based on a combination of the REBMIX and EM algorithms. The REBMIX algorithm estimates the approximate initial parameters of the Gaussian mixture model based on the known empirical probability density distribution for a given sample of observations, which are then improved using the EM algorithm. To estimate the empirical probability density required for the REBMIX algorithm, we presented several possible solutions. This resulted in three different REBMIX and EM strategies, namely Exhaustive REBMIX & EM strategy, Best REBMIX & EM strategy, and Single REBMIX & EM. The histogram is used to estimate the empirical probability density. To estimate the optimal histogram, we have developed our own optimization algorithm based on the coordinate descent algorithm. Knuth's rule was chosen to evaluate the performance of the histogram. For classification, we have developed an additional three-step procedure to philtre out irrelevant features. At the same time, we have shown a way to improve the classification performance by influencing the smoothing parameter in the histogram preprocessing of the REBMIX algorithm.
|
