Modelagem nebulosa evolutiva: novas topologias e algoritmos de aprendizagem

Detalhes bibliográficos
Ano de defesa: 2011
Autor(a) principal: Andre Paim Lemos
Orientador(a): Não Informado pela instituição
Banca de defesa: Não Informado pela instituição
Tipo de documento: Tese
Tipo de acesso: Acesso aberto
Idioma: por
Instituição de defesa: Universidade Federal de Minas Gerais
UFMG
Programa de Pós-Graduação: Não Informado pela instituição
Departamento: Não Informado pela instituição
País: Não Informado pela instituição
Palavras-chave em Português:
Link de acesso: http://hdl.handle.net/1843/BUOS-8FAH53
Resumo: This works aims to introduce new evolving fuzzy topologies and learning algorithms. Evolving fuzzy systems are defined as new class of intelligent fuzzy systems, with a high flexibility and autonomy. These systems are able to address problems like nonlinear system identification, control and pattern classification in a dynamic changing environment, adapting its parameters and structure based on a data stream. This work proposes two alternative evolving fuzzy modelling techniques. The first technique uses multivariate Gaussian membership functions defined by a non supervised recursive clustering algorithm based on participatory learning. Participatory learning is a learning model based on the human learning with a essential characteristic of robustness, since a new observation impact in causing learning or belief revision depends on its compatibility with the current system belief. Next, a novel approach for evolving fuzzy modeling is proposed, using fuzzy linear regression trees built from a stream of data in an incremental manner. The learning algorithm proposed grows the trees from a stream of data using a statistical model selection test. This model selection test is done recursively and grows the tree replacing leaves with subtrees that improves the model quality. To evaluate the model quality, the statistical model selection test takes into account the accuracy and number of the parameters of the resulting model, generating highly efficient models and avoiding over fitting. Experiments considering nonlinear system identification, time series forecasting, feature selection and fault detection and diagnosis are performed to evaluate the evolving systems proposed in this work. The results suggests these models as a promising approach for adaptive system modeling or to be used on large datasets, where the application of traditional approaches would lead to a excessive omputational cost.