Practical Application of the Original Method for Artificial Neural Network’s Training

Abstract

The article describes practical application of the original method of training artificial neural network based on the poorly formalized expert knowledge. The method allows extending the range of problems to be solved in case of lack of a sufficient number of the observations due to the fact that the training vectors are formed on the basis of the expert knowledge. The expert continuously defines classes of objects that are generated by the pseudorandom number of the training vectors of input signals, and created visual images by computer for clearly describing objects by given training vectors. The method is applied to solve important practical problem for determining of the atmospheric surface layer stability. The problem is formulated as a classification problem. As being the artificial neural network was selected multilayer perceptron. This trained neural network is represented by programming model implementing as DLL-module of dynamic-link library. The research bases on the original computer program that implements the algorithm of author’s training method. The program determines and implements the steps of the author’s research using heuristic training method of the artificial neural network to solve the problems of classification on the basis of poorly formalized experts’ knowledge. Its algorithms are used to generate visual (cognitive) images of possible situations to retrieve the unconscious expert knowledge. As a result, the aim of the study was achieved. The proposed method of training artificial neural network was applied successfully to solve a practical problem and showed its efficiency on an example of the classification problem. The author’s training method is protected by Russian patent for invention; the use of computer software holds a certificate of state registration.