Vynokurova O. Nonstationary sequences forecasting and emulation based on artificial wavelet neural networks

The dissertation work is devoted to developing intellectual methods of nonstationary sequences forecasting and emulation under condition of a priori and current uncertainty in the real time using artificial hybrid wavelet neural networks. Architecture of a hybrid wavelet neural network is modified. An adaptive modification of a learning method on the basis of stochastic approximation and a learning method of simultaneous action are proposed. For the first time generators of analytic odd and even wavelets are proposed. They enable to obtain different type of wavelets and a possibility to tune their parameters during the neural network learning process. Architecture of wavelet neuron for solving the forecasting task is developed. A new optimal on processing speed learning method of wavelet neuron using gradient methods was proposed. For the first time a learning method on the turning-points of wavelet neuron is proposed, which based on optimization of a hybrid criteria of quality. A method on the basis of evaluation planning algorithm is proposed. It has as well a protection for a sticking in the local minimum by means of input of random walking in combination with a self learning process. Simulation of developed structures and methods of learning hybrid wavelet neural models are implemented into practice.