Adaptive design of experiment for classification tree construction

Abstract

We consider pattern recognition problem under the condition that active design of experiments is available. The algorithm of adaptive construction of classification tree is suggested. In the consequent design of experiment, the algorithm takes into account logical regularities with correspondent estimates of risk obtained on previous stages of the analysis. For risk evaluation, we use the Bayes model of recognition on a finite set of events. The results of statistical modeling with small number of observations and heterogeneous features demonstrate the effectiveness of the proposed methodology.