Genetic Feature Selection for Fuzzy Discretized Data

Luciano Sánchez, José R. Villar, Inés Couso

A wrapper-type evolutionary feature selection algorithm, able to use fuzzy data, is proposed. In the context of Genetic Learning of Fuzzy Rule-based Classifier Systems (FRBCS), this new algorithm has been applied to a particular kind of instances, comprising fuzzy discretized data (FDD). This data is obtained when passing crisp data through the fuzzification interface of the FRBCS under study. We have compared the properties of the algorithm proposed here to other approaches, over FDD and crisp data. In case the preprocessed data is intended to be used by a Genetic Learning FRBCS, we can conclude that those algorithms able to use FDD are preferred over the crisp ones, even though there is not fuzziness in the training data being used. Besides, they also are the only alternative when the datasets are imprecise, although this last case is not elaborated in this study.

PDF full paper