Self-organization of probabilistic network with Gaussian mixture model

A new method of constructing probabilistic networks based on the self-organization of a Gaussian mixture model is presented. The distinctive features of the proposed self-organization method are as follows: (1) it is assumed that a class is composed of a number of subclasses of Gaussian PDFs or Gaussian kernels, such that the given class samples are decomposed probabilistically into subclass samples based on the current subclass PDFs. The discrepancy between the current and the actual subclass PDFs represented by the decomposed subclass samples invokes iterative update of the current subclass PDFs until an equilibrium is reached. (2) The required number of subclasses or Gaussian kernels are automatically recruited as necessary for avoiding local minima encountered during the iterative update of subclass PDFs. The detection of local minima is done by applying the Chi-square test to individual subclasses. The proposed method provides the semi-parametric estimation of an arbitrary form of PDFs, which allows the network to avoid local minima and obtain the maximum likelihood estimate. Simulation results show the capability of the proposed method for accurately estimating class PDFs