Abstract: In many real classification problems where a limited number of training samples is available, the linear classifiers based on discriminant analysis are unable to deliver accurate results.