Control and Cybernetics

A new system for selection of reference instances, which is called the EkP system (Exactly k Prototypes), has been introduced by us recently. In this paper we study suitability of the EkP method for training data reduction on seventeen datasets. As the underlaying classifier the well known IB1 system (1-Nearest Neighbor classifier) has been chosen. We compare generalization ability of our method to performance of IB1 trained on the entire training data and performance of LVQ, Learning Vector Quantization, for which the same number of codebooks has been chosen as the number of prototypes selected by the EkP system. The comparison indicates that even with only a few prototypes which have been chosen by the EkP method on nearly all seventeen datasets statistically indistinguishable results from those given by the IB1 system have been obtained. On many datasets generalization ability of the EkP method has been higher than the one attained with LVQ.