Boiko O. Evolving neuro-fuzzy systems in data mining tasks

The thesis is devoted to development of evolving neuro-fuzzy systems that are able to tune not only their synaptic weights and membership functions parameters, but also their architecture in the online mode. Known neuro-fuzzy systems architectures and their learning methods are analyzed. The hybrid learning method of evolving multilayer neuro-fuzzy system based on the Wang-Mendel system is proposed. This learning method combines architecture evolving processes, membership functions self-learning and synaptic weights learning. Learning methods for all parameters of the neuro-fuzzy nodes tuning are proposed. These methods improve approximation capabilities of the evolving neuro-fuzzy systems. Learning methods for evolving systems based on the GMDH and the cascade systems are proposed. These systems use two-input Wang-Mendel neuro-fuzzy nodes and two-input neo-fuzzy nodes. The architecture and learning methods for weighted ANARX-model are proposed. The WANARX-system is used for non-stationary nonlinear time series forecasting. The evolving neuro-fuzzy Kohonen network and its learning method are improved. This system can process data in the online mode without prior knowledge about the number of clusters.