Navegando por Data de Publicação, começando com "2015-09-17"

Filtrar resultados por ano ou mês

Agora exibindo 1 - 1 de 1

Uma abordagem para construção de sistemas fuzzy baseados em regras integrando conhecimento de especialistas e extraído de dados
(Universidade Federal de São Carlos, 2015-09-17) Lima, Helano Póvoas de; Camargo, Heloisa de Arruda; https://lattes.cnpq.br/0487231065057783; https://lattes.cnpq.br/4850820391832869
Historically, since Mamdani proposed his model of fuzzy rule-based system, a lot has changed in the construction process of this type of models. For a long time, the research efforts were directed towards the automatic construction of accurate models starting from data, making fuzzy systems almost mere function approximators. Realizing that this approach escaped from the original concept of fuzzy theory, more recently, researchers attention focused on the automatic construction of more interpretable models. However, such models, although interpretable, might not make sense to the expert. This work proposes an interactive methodology for constructing fuzzy rule-based systems, which aims to integrate the knowledge extracted from experts and induced from data, hoping to contribute to the solution of the mentioned problem. The approach consists of six steps. Feature selection, fuzzy partitions definition, expert rule base definition, genetic learning of rule base, rule bases conciliation and genetic optimization of fuzzy partitions. The optimization and learning steps used multiobjective genetic algorithms with custom operators for each task. A software tool was implemented to support the application of the approach, offering graphical and command line interfaces and a software library. The efficiency of the approach was evaluated by a case study where a fuzzy rule-based system was constructed in order to offer support to the evaluation of reproductive fitness of Nelore bulls. The result was compared to fully manual and fully automatic construction methodologies, the accuracy was also compared to classical algorithms for classification.