“Feature Selection Based on Genetic Algorithms that Achieves High Recognition Rates and Low Numbers of Dimensions and its Application to Character Recognition”

In this paper, we propose an improved feature selection method based on genetic algorithms (GA) that achieves both high recognition rates and low numbers of dimensions. This method performs the first search using a GA that uses recognition rate as fitness, and performs the second search with a GA which uses the number of dimensions (reduction) as fitness while avoiding a drop in recognition rate. To speed up the second search,we introduce a special chromosome with maximum reduction quantity that treats all features as unselected as the information source of the GA. We perform experiments using five sets of hand-written Kanji character patterns. Each set consists of patterns in two similar categories. The results show that the proposed method reduces the number of dimensions to 6.2% of the original value wile providing higher recognition rates than the method that uses all features.