You are here

Neuroevolutionary methods of information processing and control

Work number - M 75 FILED

Bezsonov O.O.

Kharkiv National University of Radio Electronics

The relevance of the cycle of scientific papers suggests that artificial neural networks (ANN ) was obtained recently becoming more common and are the theoretical basis of neuro-computers. ANN convincingly proven effective in solving various problems of adaptive information processing , such as a nonlinear function approximation , classification, control, signal and image filtering , etc., which in one form or another is used approximation of complex nonlinear dependencies. But despite all these advantages, the INS and currently are very limited practical use in application systems. In this paper, the author proposes the structure of a network by solving multiobjective optimization (Pareto optimization) and implement its adaptation using evolutionary techniques, in particular the genetic algorithm, which can significantly expand the use of ANN in solving practical problems.

Following research results are obtained:

1. First developed multistep projection methods of training neural networks, which, unlike the method of least squares, require significantly less information required for network training and developed their recurrent form, which provides a significant reduction in the learning process.

2. First proposed the use of asymmetric functional quality and are designed based on these algorithms ANN training that allows you to receive high-quality model in non-gauss interference and improve the efficiency of constructing econometric models of inventory management and security etc.

3. Improved model of computational tools that implement hierarchical neural networks in problems of compression and image filtering, which significantly reduce the hardware cost and memory capacity to store images, and reduce processing times.

4. Further developed an evolutionary approach to determine the optimal structure of adaptive neural networks and their parameters, which increases the efficiency of the INS when they are used to optimize the transient objects in the absence of information about the properties of the medium in which the facility operates.

The practical significance of the work lies in the fact that the developed neural network models, methods and structures that allow us to solve the problem of identification and control of nonlinear dynamic objects, filtering of signals and images in real time.
Total number of publications of the author - 79, including 30 articles in professional journals. Number of publications in refereed database "Ukraїnіka Naukova" - 16, in the database SCOPUS - 8, according to the citation index SCOPUS - 2.