ファジイ帰納学習アルゴリズムの改良

抄録

Inductive learning algorithms acquire knowledge from a training example set. The algorithms basically deal with discrete attributes. To deal with numerical attributes, the algorithms must have the procedure which transforms numerical values to discrete values. We proposed a fuzzy inductive learning algorithm IDF which learned a fuzzy decision tree from a training example set. IDF also has the procedure which transforms an attribute range to fuzzy ranges and can deal with numerical and fuzzy values. However, this procedure tends to decompose the attribute range for the upper attribute in the fuzzy decision tree to many fuzzy ranges. Thus we propose new procedure which composes redundant fuzzy ranges using Akaike's information criterion and improve the IDF. We made numerical experiments for some real training example sets and verified the efficiency of new IDF. We also compare new IDF with C4.5 which is a representative inductive learning algorithm and show that IDF is more efficient than C4.5 in terms of classification accuracy.