Gorshkov Y. Data classification under uncertainty conditions on the basis of hybrid neuro-fuzzy architectures

The thesis is devoted to a research of the hybrid neuro-fuzzy network architectures, learning and self-organization methods, adaptive fuzzy clustering procedures in the problem of data classification under a priory and current uncertainty conditions with respect to the type of data distribution and vastly overlapping classes. Robust probabilistic and possi-bilistic recursive fuzzy clustering methods which allow clas-sification in the conditions of substantial class overlapping; the modified probabilistic neural network with fuzzy infer-ence, network learning and growing methods; recursive learning method based on ellipsoidal approach for the radial-basis neural networks, which provides effective learning un-der the uncertainty conditions about the noise; hybrid neuro-fuzzy counter-propagation network and its learning method are proposed for the first time. Self-organization methods of the Kohonen network with fuzzy inference are improved and the data preprocessing method by increasing of the input space dimensionality is proposed. Experimental results con-firm the efficiency of the proposed methods on a number of well-known benchmarks and the real fuzzy data classification problems.