Initializing EM algorithm for univariate Gaussian, multi-component, heteroscedastic mixture models by dynamic programming partitions

摘要

Setting initial values of parameters of mixture distributions estimated byusing the EM recursive algorithm is very important to the overall quality ofestimation. None of the existing methods is suitable for mixtures with largenumber of components. We present a relevant methodology of estimating initialvalues of parameters of univariate, heteroscedastic Gaussian mixtures, on thebasis of the dynamic programming algorithm for partitioning the range ofobservations into bins. We evaluate variants of dynamic programming methodcorresponding to different scoring functions for partitioning. For simulatedand real datasets we demonstrate superior efficiency of the proposed methodcompared to existing techniques.