This paper is devoted to problems of Learning theory as well as Machine Learning 
and pattern recognition. The target of the paper is  to estimate the generalized 
ability of the algorithm, using information from data as much as possible. 
To realize this idea the approach consisted of two parts has been proposed. 
One of them is combinatorial approach for estimating the probability of correct recognition. 
And the second one is probabilistic approach that allows us to estimate 
the result of internal structure of the algorithm. Both of them give more precise estimate 
of the general probability of correct recognition.