A new method for generating statistical classifiers assuming linear mixtures of Gaussian densities

Introduces a new method for generating Bayes classifiers assuming linear mixtures of Gaussian probability densities. This new classifier adapts to the data set, finding and using the minimum number of Gaussian probability densities needed to discriminate between classes. In brief the concept is to first design Bayes classifiers assuming Gaussian densities. Next, if the error rate is unacceptable, the number of Gaussian densities in (the mixture distribution of) one of the classes is increased by one, new classifier parameters are estimated and the (new) error rate is computed. This process of classifier generation and evaluation continues until a set of criteria is fulfilled. Finally, one of the generated classifiers is selected. Comparisons with other relevant classifiers, using both synthetic and real data sets, show that the author's method generates reliable classifiers.