Abstract
I proposed an fuzzy inductive learning algorithm IDF, where IDF generates rules in form of fuzzy decision tree. I verified the effect of IDF by numerical experiments. However, it is necessary for IDF to collect enough training samples. On the other hand, it is difficult to collect enough samples in real world problems. Because we don't know indispensable samples for a inductive learning algorithm and we can't collect samples that include all combination of attribute values. Thus, it is necessary to generate rules from the collected samples. The training samples are usually insufficient to learn all rules. After we generate the rules once, it is necessary to collect new samples and refine the rules. Thus, the paper proposes new approach to collect necessary samples and refine rules in IDF. The approach identifies evaluation samples with bad evaluation accuracy and collect the samples as training samples. The approach refines the rules with new samples. The paper also verifies the effect. of new approach by munerical experiments with a real world problem.
