Abstract
A broad class of pattern recognition problems deals with a direct classification task. It is based on measuring a distance between a pattern to be classified and some prototypes of classes studied in the problem. Quite frequently these prototypes are not provided and have to be computed. A way of building prototypes of classes one is interested in, not necessarily only those directly available in the training (learning) set, is studied. The algorithm is aided by a referential neural network structure. Also introduced are some self-flagging mechanisms determining the feasibility of calculated prototypes. The computational framework is constructed in terms of neural networks. The conceptual knowledge representation counterpart of the classification problem is developed making use of selected concepts of fuzzy sets.