Informatica

Let G0 and G1 be arbitrary fuzzy classifiers (Vatlin, 1993). We say that G1 improves G0 if the performance of G1 is more than G0 one. We also introduced the concepts of consistent and strongly selfguessing fuzzy classifiers. The criterion of strong selfguessing is formulated. The theorems on the conditions of probabilistic improvement of consistent and monotonic improvement of strongly selfguessing fuzzy classifiers are proved.

The problem of construction of the fuzzy classification models (fuzzy classifiers) with high generalization ability is discussed. The strong self guessing property of fuzzy classificational models is introduced and examined. It is proved that this characteristic doesn't form a full system of restrictions, i.e., for the unambiguous detection of the most valid fuzzy classifier (among the set of fuzzy classifiers agreed with arbitrary learning set) it is necessary to use additional “regularizing” restrictions.