FUZZY CLASSIFIER LEARNING BASED ON DISTANCE BETWEEN THE MAIN COMPETITORS

Abstract

The classification problem is the assignment an object with certain features to one of classes. Various engineering, management, economic, political, medical, sport, and other problems are reduced to classification. In fuzzy classifiers «inputs – output» relation is described by linguistic rules. Antecedents of these rules contain fuzzy terms «low», «average», «high» etc. To increase the correctness it is necessary to tune the fuzzy classifier on experimental data. The new criteria for fuzzy classifier learning that take into account the difference of membership degrees to the main competitors only are proposed. When the classification is correct, the main competitor of the decision is the class with the second largest membership degree. In cases of misclassification the wrong decision is the main competitor to the correct class. Computer experiments with learning the fuzzy classifier of 3 kinds of Italian wines recognition showed a significant advantage of the new criteria. Among new learning criteria the criterion in the form of squared distance between main competitors with the penalty for wrong decision has minor advantage. New criteria can be used not only for tuning fuzzy classifiers but for tuning some other models, such as neural networks.