Research object - dynamic stochastic processes under conditions of complete or partial, a priori and current uncertainty about the mathematical models of such processes and their parameters. Research target is the synthesis of hybrid artificial evolving neuro-fuzzy networks and learning algorithms for them with increased performance, capability to provide linguistic interpretation of the outputs, and possibility to adjust their structure during data processing in sequential mode. Methods of research: theory of artificial neural networks - to synthesize new growing architectures of neural networks; fuzzy logic - to made possible to produce fuzzy inference out of the suggested architectures; optimization theory - for synthesis of new learning algorithms with increased performance and possibility to process noisy data; apparatus of mathematical statistics - to analyze received results and make scientific and practical conclusions. Theoretical and practical results of the thesis in general solve scientific problems of forecasting and classification under the conditions of high-level uncertainty using the hybrid evolving neural networks. Scientific novelty: 1) specialized architectures of ortho-synapse, ortho-neuron and double-neuron, which use classical systems of orthogonal polynomials as activation functions, are proposed for the first time, as well as methods for adjustment of their synaptic weight coefficients in batch mode and mode of sequential information processing; 2) multi-dimensional architecture of the cascade neo-fuzzy neural network is proposed for the first time, which is a neuro-fuzzy system with multi-layered fuzzy inference, capable of processing multi-dimensional input and output data faster than classical architectures, and with capability to adjust its own architecture automatically, adapting to changes in external factors of processed data; 3) learning methods of cascade neural networks based on exponentially weighted recursive least squares Peterka method, Greville theorem and Frobenius formula for inversion of matrixes with a large dimensions are proposed for the first time, allowing to solve problem in sequential mode of information processing and making it possible to increase performance of suggested neuro architecutres in comparison to conventional neural networks; 4) modified cascade-correlation Fahlman & Lebiere architecture by replacing the artificial neurons in the nodes of architecture with ortho-neurons, quadratic neurons, and neo-fuzzy neurons, what lead to significant reduction of training time in comparison with the prototype, a capability to obtain linguistic interpretation of the output signals, and simplifying the architecture for the implementation in hardware, according to the selected type of artificial neuron in the nodes; 5) method of self-organizing neural network architecture based on the GMDH is improved by replacement of N-Adalines used conventionally with the neo-fuzzy neurons, what allows automatically receive neuro-fuzzy architecture with optimal complexity, which provides a linguistic interpretation of the output signal by means of multilayered fuzzy inference. Degree of implementation - the research results are implemented in the State Scientific-Production Enterprise "Systemni Tekhnolohii" (act of 08/06/2010), scientific statements, conclusions and recommendations contained in the thesis, were used for preparation of the "Neural network methods of computational intelligence" course, which is taught to students of the specialty "Intelligent Decision Support Systems" at the Kharkov National University of Radio Electronics (act of 09/15/2010), as well as in research projects of Kharkov National University of Radio Electronics (act of 5/26/2010). The scope of use - in organizations that deal with problems of intelligent data processing and in the areas of finance, energy, petrochemicals, transportation, medicine, biology and ecology in the learning process in the preparation of specialists in the field of intellectual processing information.
