Bezsonov O. Evolving artificial feedforward neural networks: architecture, training, applications

The thesis analyzes the problems of intellectual data mining theoretical foundations and creation of new evolutionary artificial neural networks (EANN) for improving the efficiency of information processing in conditions of priori and current uncertainty. A new method of robust multi-objective optimization (Pareto optimization) based on robust fitness functions and information criteria for assessing the complexity of the model, is developed. New methods of training EANN that ensure the required accuracy in the presence of limited symmetric and asymmetric noise are proposed. New procedures for the functional parameters correction and evaluating parameters of noise that is described by Tukey-Huber model are developed. New laws of adaptive neuro-predictive control of nonlinear non-stationary dynamic objects that operate under uncertainties are designed. Simulation of different EANN training algorithms was performed in NeurophStudio environment. The process of identifying and solving problems of nonlinear dynamic objects predictive control is researched. Software tools that implement the proposed methods for constructing EANN are developed. Experimental research of developed methods properties and characteristics, that confirmed the fundamentals of the thesis, is performed. The validity of the obtained results is confirmed by the experimental studies and real applications.