Agrupamento Nebuloso Evolutivo Baseado em Critério Gaussiano Multivariável para Tomada de Decisão em Fluxos de Dados

Abstract

In this dissertation, an evolving fuzzy clustering algorithm applied to data streams is proposed. The algorithm performs the clustering of the data using a “from scratch” approach, i.e., determining its variables completely from the streaming data, starting from the rst data sample. It is based on computing the potential of the data to make decisions about the centers of clusters at each sample, using a multivariable Gaussian function that considers the density and the rate of variation as information variables. The algorithm includes an autonomous data normalization mechanism for similarity in scaling according to the dynamic operating range of the data stream and constructs clusters with timevarying prototypes based on fuzzy covariance matrices and an exponential distance metric. In its decision-making process, it employs a mechanism based on clusters’ sensitivities and adaptive cluster radii. For updating and merging clusters, the algorithm includes a mechanism based on the evolutionary crossover technique. Computational results consider the implementation of the algorithm for evolving fuzzy clustering of 8 datasets benchmarks (S1, A1, R15, Unbalance, Gauss 6-dim, Gauss 10-dim, DIM032, and DIM064) widely used in data stream clustering tasks, as well as for evolving fuzzy clustering of data applied to online time-series modeling (Hyperbole Function, Mackey-Glass, and Nakanishi) and to online identication of nonlinear dynamic systems (SISO and MIMO). Experimental results consider the implementation of the algorithm for evolving fuzzy clustering of data applied to online identication of real-world dynamic systems (Thermal Plant, 2DoF Helicopter, and Fogtrein-I Rocket). In the online time-series modeling and online dynamic systems identication processes, the algorithm is applied in the fuzzication of data and in the creation/updating of evolving fuzzy rules in an evolving fuzzy inference system. For the consequent proposition of the evolving fuzzy rules, the state-space model is adopted, with its parametric identication performed using the Recursive Fuzzy Observer/Kalman lter Identication/Eigensystem Realization Algorithm (OKID/ERA). In general, the results indicate that the algorithm is suitable for evolving fuzzy clustering of data streams, achieving competitive performance in comparison to other algorithms in the literature.