Slipchenko O. Artificial neural networks with variable number of nodes in information processing problems

The dissertation deals with the development of intelligent methods for nonstationary sequences forecasting and emula-tion under conditions of a priori and current uncertainty in real time using artificial hybrid neural networks with variable number of nodes. Architecture of a hybrid neural network for solving forecasting and identification problems is modified. The use of orthogonal activation function allows obtaining neural networks with a number of useful properties. Modifi-cation of the least-squares method for neural networks learn-ing is proposed. The modified method uses triangular weight-ing function which allows reduced identification delay com-pared to known methods. For the first time, a method for modifying structure of a neural network without requiring its retraining after adding or removing a node is proposed. A new optimal method for training an ensemble of neuropredic-tors is proposed. A new method for modifying structure of an ensemble of predicting neural networks is also proposed. Simulation of the developed structures and learning methods for hybrid neural models and ensembles of neural networks are implemented in practice.