LEARNING FINITE GAUSSIAN MIXTURES USING DIFFERENTIAL EVOLUTION
Wojciech Kwedlo · 2010
In the paper the problem of parameter estimation of finite mix ture of multivariate Gaussian distributions is considered. A new approach based on differential evolution (DE) algorithm is proposed. In order to avoid problems with infeasibility of chromosomes our version of DE uses a novel representation, in which covariance matrices are encoded using their Cholesky decomposition. Numerical experiments involved three version of DE differing by the method of selection of strategy parameters. The results of experiments, performed on two synthetic and one real dataset indicate, that our method is able to correctly identify the parameters of the mixture model. The method is also able to obtain better solutions than the classical EM algorithm.