The problem of selection the most informative features is reduced to an optimization
problem for the average risk functional whose maximization is equivalent to maximization of
informational distance between distributions of features in two classes. We consider a maximization procedure for the average risk functional via empirical risk, estimating the divergence
between them, with Rademacher complexity. The proposed method has been applied efficiently
to problems of selection parameters important to separate the states of technological processes.
We show an experimental comparison of the developed approach with other widely known
feature selection techniques.