Kolchygin B. Adaptive neuro-fuzzy systems for fuzzy cluster analysis under conditions of uncertainty

3. The thesis is devoted to research and development of adaptive fuzzy clustering systems using Kohonen neural networks and their ensembles. Prototype-based clustering methods and their modifications are considered. It is noticed that as the methods of unsupervised learning, any clustering methods have some a priori view about the nature of the data distribution in the treated sample, and work successfully only if data meets these expectations. Building a clustering system, successfully operating in unknown in advance conditions or in cases of changing characteristics of the data during work is possible only with using of methods of collective reasoning. It involves the synthesis of the current model in on-line mode based on parallel processing systems, each of which is successful only on some part of the sample. The construction of such a system on the basis of clustering systems fraught with difficulties due to the formulation of the problem of clustering, primarily the lack of objective criteria for estimating the quality of the partition, as well as a large number of adjustable parameters and using mainly batch methods to process the data. The thesis proposes a number of clustering methods, covering the most common methods of processing the sources of nonstationarity in data and modified to be able to work in on-line mode. All methods are based on the common approach to minimize the number of adjustable parameters, and give them a clear physical meaning, as well as providing opportunity to work with a growing sample. Thus obtained clustering methods work in some sense uniformly, that allows to combine them into neural network ensembles for the collective result of the partitioning matrix, the best of each of those obtained by individual methods. Using the features of the fuzzy prototype-based clustering methods, the clustering system in the paradigm of fuzzy logic type-2 working completely in on-line mode was implemented first. Using such a powerful tool has demonstrated its effectiveness in the growing data samples in which there is a significant properties' drift over time: the appearance and disappearance of clusters, changing their characteristic scale and density of outliers, the degree of overlapping, etc. An experimental study of the properties and characteristics of the developed methods is carried out and recommendations on their use in solving practical problems are proposed.