In this paper the application of an instance selection algorithm to the design of a neural classifier is considered. A number of existing instance selection methods are presented. A new wrapper-method, whose main difference compared to other approaches is an iterative procedure for selecting training subsets from the dataset, is described. The approach is based on using training subsample selection probabilities for every instance. The value of these probabilities depends on the classification success for each measurement. An evolutionary algorithm for the design of a neural classifier is presented, which was used to test the efficiency of the presented approach. The described approach has been implemented and tested on a set of classification problems. The testing has shown that the presented algorithm allows the computational complexity to be decreased and the quality of the obtained classifiers to be increased. Compared to analogues found in scientific literature, it was shown that the presented algorithm is an effective tool for classification problem solving.