Structure Identification of Fuzzy Classifers

For complex and high-dimensional problems, data-driven identification of classifiers has to deal with structural issues like the selection of the relevant features and effective initial partition of the input domain. Therefore, the identification of fuzzy classifiers is a challenging topic. Decision-tree (DT) generation algorithms are effective in feature selection and extraction of crisp classification rules, hence they can be used for the initialization of fuzzy systems. Because fuzzy classifiers have much flexible decision boundaries than DTs, fuzzy models can be more parsimonious than DTs. Hence, to get compact, easily interpretable and transparent classification system, a new structure identification algorithm is proposed, where genetic algorithm (GA) based parameter optimization of the DT initialized fuzzy sets is combined with similarity based rule base simplification algorithms. The performance of the approach is studied on a specially designed artificial data. An application to the Cancer classification problem is also shown.