Open AccessHYBRID DECIDING MODULES WITH VIRTUAL STREAMS FOR CLASSIFICATION AND PREDICTION OF FUNCTIONAL STATE OF COMPLEX SYSTEMS
Author(s) -
Andrey Kiselev,
T. V. Petrova,
S. V. Degtyaryov,
A. F. Rybochkin,
С. А. Филист,
О. В. Шаталова,
В. Н. Мишустин
Publication year - 2018
Publication title -
izvestiâ ûgo-zapadnogo gosudarstvennogo universiteta
Language(s) - English
Resource type - Journals
eISSN - 2686-6757
pISSN - 2223-1560
DOI - 10.21869/2223-1560-2018-22-4-123-134
Subject(s) - computer science , artificial neural network , data mining , artificial intelligence , nonlinear system , machine learning , dimension (graph theory) , basis (linear algebra) , support vector machine , mathematics , physics , geometry , quantum mechanics , pure mathematics
The problem reviewed of building intelligent decision support systems for classification and prediction of the functional state of complex systems in the article. To predict the state of complex systems, hybrid decision modules with virtual flows are proposed, which reflect the hidden system connections between real and virtual data. The vector of informative features at the input of the hybrid decision module consists of two subsectors, the first of which corresponds to real flows, and the second - to virtual flows. Simulation modeling of classification processes using latent variables was performed, which allowed to evaluate the effect on the quality of classification of artificially introduced virtual flows. The structure of a neural network model with virtual recurrent-type streams is developed. The structure consists of N consecutively included neural network approximants. The outputs of the previous approximators are combined with the vector of in-formative attributes of the subsequent approximators, which allows forming virtual flows of different dimensions. A method is developed for the formation of non-linear models of virtual flows, characterized by the use of the GMDH-simulation method to obtain models of the influence of real flows on virtual flows, learned through nonlinear adalines. The method makes it possible to form a subvector of latent variables of unlimited dimension. Non-linear models of virtual flows are formed through a method based on the use of GMDH modeling. The method makes it possible to obtain neural network structures built on the basis of GMDH models and nonlinear adalines, which make it possible to form a subvector of latent variables of unlimited dimensionality.