LEARNING OF SOFT CLASSIFIER VIA DIFFERENTIAL EVOLUTION

The study is oriented to two basic aims: define and then realize a classfier with constrained gain and try to learn it to have maximum possible sensitivity. Their meaning is to have maximum possible quality of classification into given number of classes together with suppressed magnification of any input details. It is necessary to define soft classfier, its gain and soft sensitivity, first. The process of soft classfier learning is then described as constrained minimization task on convex domain. But the objective function is non-smooth and non-convex in general, which makes the learning non-trivial and difficult, unfortunately. There is a good chance for sophisticated heuristics of global minimization. One of them, which is based on the competition of various procedures of differential evolution, was used to learn multiclassifier on famous and traditional example of Fishers iris flower classification. The paper is also about the global minimization of non-smooth functions and its role in the domain of artificial intelligence.