Vynokurova O. Hybrid evolving adaptive wavelet-neuro-fuzzy systems in dynamic data mining

The thesis is dedicated to solving a topical problem of dynamic data mining methods design for on-line non-stationary nonlinear signal processing based on hybrid evolving adaptive wavelet-neuro-fuzzy systems, which can operate under priori and current lack of information conditions. Such methods are characterized by improving learning rate and have possibility of processing the time series with short and long data sampling, as well as with local features and abnormal outliers with unknown distribution. For the first time a number of wavelet-neural networks, hybrid evolving adaptive wavelet-neuro-fuzzy systems, type-2 hybrid wavelet-neuro-fuzzy systems, intelligent adaptive control laws based on proposed wavelet-neuro-fuzzy models, and hybrid evolving multirowed and cascaded GMDH-wavelet-neural network for the dynamic data mining tasks solving are proposed. For the suggested architectures of hybrid evolving adaptive wavelet-neuro-fuzzy-systems a number of learning methods based on quasi-Newtonian and robust procedures are proposed. These learning methods are characterized by increased learning rate, following and filtering properties.