The modified method of a nonparametric estimation of a multidimensional probability density of features at the limited possibilities of the classifier

The nonparametric estimation of a multidimensional probability density of features on the set learning sample of their vectors in the conditions of the limited memory size and the limited speed of computing possibilities of the recognition device is necessary for solving many problems of statistical recognition. In the article, the method of such estimation on the basis of Parzen windows (contributions) by expansion ones in series is presented.

The method differs from known techniques, that those contributions are subject to decomposition only, which are included into groups (pairs, triples, etc.) the nearest on the chosen measure of adjacency. These measures of adjacency are selected so, that after generalization of each of groups of contributions and groups of training vectors corresponding to them in the generalized contribution and to the generalized vector demanded restriction of an estimation length and minimization of an error arising because of this restriction is provided. The modified method of the normal contributions providing the aforementioned solution, is described in conditions when the pairs of contributions are used as groups.