In this paper a multiclass classification problem solving technique based on genetic programming is presented. Classification algorithms are designed to learn a function which maps a vector of object features into one of several classes; this is done by analyzing a set of input-output examples of the function (also called "training samples"). Here we present a method based on the theory of genetic algorithms and genetic programming that interprets classification problems as optimization problems. The major aspects presented in this paper are suitable genetic operators for this problem class (mainly the creation of new hypotheses by merging already existing ones and their detailed evaluation) we have designed and implemented. We define a novel function for measuring a classificator model's quality that takes into account several different features of the model to be evaluated; an extended version of ROC curves that can be applied not only to two-class-classification but also to multiclass classification problems, is also presented. The experimental part of the paper documents the ability of this method to yield very satisfying results; selected results achieved for two classification benchmark problems are discussed.