Bayes Classification for Nonstationary Patterns

The paper's subject is classification with nonstationary patterns. The attribute space is finite-dimensional, while its coordinates in particular may be continuous, binary, discrete, categorical in character, or also a combination of these. The number of patterns is not methodologically limited. Use of the Bayes approach minimizes the expected value of misclassifications, allowing additionally for an influence in the proportions of probability of errors when assigning to specific classes. In turn, the statistical kernel estimators method makes the algorithm independent of patterns' shapes. The investigated procedure also eliminates elements of patterns which have insignificant or even negative influence on the results' accuracy. Appropriate modifications follow the classifier parameters, which increases the effectiveness of procedure adaptation for nonstationary patterns. The algorithm concept is based on the sensitivity method, used with artificial neural networks.