Variants of heuristic rule generation from multiple patterns in Michigan-style fuzzy genetics-based machine learning
Yusuke Nojima, Kazuhiro Watanabe, Hisao Ishibuchi · 2015
In the design of rule-based classifiers, a single rule is often generated from a single pattern in a heuristic manner. Since the generated rule is likely to be over-specialized to the pattern, its conditions are often randomly replaced with don't care. However, the generalized rule with don't care conditions does not always have high classification ability. This is because the replacement is randomly performed without utilizing any information about other patterns. In our previous studies, we proposed an idea of generating a fuzzy classification rule from multiple patterns. In this paper, we propose its six variants. Each variant has a different criterion for choosing multiple patterns from which a single rule is generated. The proposed variants are used to generate fuzzy classification rules in Michigan-style fuzzy genetics-based machine learning. The usefulness of each variant is evaluated as a heuristic fuzzy rule generation method through computational experiments on 20 benchmark data sets.