Kondratiuk O. Methods of analysis and synthesis of activation functions by designing of neural networks for time series forecasting.

The thesis is devoted to an increasing of property (accuracy) of neural network forecasting of time series by development and using methods of analysis and synthesis of activation functions of neural networks on the base of purposeful control their sensitivity. In work concept of sensitivity of neural networks was entered and quantity indicator for its estimation was developed. It is offered the development of structural approach to improvement of property of neural network forecasting on the base of preliminary analysis and account of statistical characteristics of input data. It is discovered dependence between the accuracy of forecasting and the measure of concordance of activation functions of the input layer neural network to the distribution input data. It is gotten the development the methods of analysis and synthesis of the activation functions neural network. The practical realization of the developed methods was executed in the form of program neuroemulator Neiro.