Prequential Bayes mixture approach for Gaussian mixture order selection

This paper presents a modified prequential Bayes (MPB) method for model order estimation of Gaussian mixture models (GMM). The proposed MPB order estimators recursively update the weighting for each order in a class of model orders from the mixture of a time-invariant prior and the likelihood of the observed data for each model. This paper investigates both a maximum a posteriori (MAP) switching version and an affine version of the MPB order estimator. Simulations demonstrate that the proposed MPB estimators are more accurate for small sample sizes than the minimum description length (MDL) criterion and the Akaike information criterion (AIC).