Advances in Computer Science Research Estimation of Parameters of Gaussian Mixture Models by a Hybrid Method Combining a Self-adaptive Differential Evolution with the Em Algorithm

In the paper the problem of learning of Gaussian mixture models (GMMs) is considered. A new approach based on hybridization of a self-adaptive version of differential evolution (DE) with the classical EM algorithm is described. In this approach, called DEEM, the EM algorithm is run until convergence to fine-tune each solution obtained by the mutation and crossover operators of DE. To avoid the problem with parameter representation and infeasible solutions we use a method in which the covariance matrices are encoded using their Cholesky factorizations. In a simulation study GMMs were used to cluster synthetic datasets differing by a degree of separation between clusters. The results of experiments indicate that DE-EM outperforms the standard multiple restart expectation-maximization algorithm (MREM). For datasets with high number of features it also outperforms the stateof-the-art random swap EM (RSEM).