Mathematics & Physics 2016, 9(4), 473-480 УДК 519.87 Selecting Informative Variables in the Identification Problem Eugene D.Mihov∗ Oleg V.Nepomnyashchiy† Institute of Space and Information Technology Siberian Federal University Kirensky, 26, Krasnoyarsk, 660074 Russia Received 23.06.2016, received in revised form 14.08.2016, accepted 14.09.2016 The problem of multidimensional object classification with small training sample is considered. <...> The following algorithms of estimating variable informativeness are considered: Ad, Del, AdDel. <...> A new algorithm for selecting informative variables is proposed. <...> It is based on the optimization of the coefficient vector of the kernel fuzziness. <...> Some modification of this algorithm is also discussed. <...> Keywords: classification, small training sample, informative variable, optimization of the coefficient vector of the kernel fuzziness. <...> Introduction Nowadays, the classification problem is solved by many ways. <...> One of the widespread classification methods uses neural networks. <...> Classification methods based on parametric model and non-parametric algorithms are also used. <...> First of all one must select object variables that are used for classification. <...> Modern classification objects can be characterized by many variables, but not all variables reflect the object membership to a particular class. <...> The variables that do not reflect the object membership to a particular class are uninformative, and those variables that reflect the object membership to a particular class are informative. <...> The modern classification objects can be characterized by many variables. <...> For example, to identify the class of the patient disease one can carry out the large number of tests, and obtain hundred or more variables characterizing the patient health, but not all variables are informative. <...> To classify an object, it is necessary to have a large training sample. <...> However, large training samples are not often available. <...> According to common rules, it is impossible to draw a conclusion on the basis of 600 sample elements if there are 20 variables." One should note that for such tasks the estimation of variable informativeness has the particular importance. <...> It is difficult to create the decision rule for large number of variables and a ∗edmihov@mail.ru †ONepomnuashy@sfu-kras.ru ⃝ Siberian Federal University. <...> All rights <...>