Genetic Feature Selection for Low Quality Data

Acquiring precise data is expensive. Hence, there is interest in the development of computer algorithms that make full use of imprecise data, even though these new algorithms may be more complex and require more resources. In particular, most Genetic Fuzzy Systems (GFS). accept crisp inputs, but changes can be effected to them so that rules are obtained from vague data. Furthermore, the rule generation is only one stage in the design of a model. If a GFS uses vague data, we also need to preprocess this low quality data before the learning can take place, but there are very few algorithms capable of selecting vague features or detecting redundant imprecise instances, for example. In this work we propose a wrappertype evolutionary feature selection algorithm, able to use incomplete and imprecise data. In the context of Genetic Learning of Fuzzy Rulebased Classifier Systems (FRBCS), we have applied it to remove unnecessary features of fuzzy discretized data.