1996 年 11 巻 4 号 p. 600-607
In this paper, a new method for knowledge acquisition is proposed. This method involves an interactive dialog between a learning engine and a human expert. Knowledge is hard for experts to recognize by themselves. They can, however, point out errors and recognize a lack of knowledge. Research in this field to date tends to improve reasoning or learning to obtain more precise knowledge. This approach, however, requires completeness of sample data and domain knowledge, which is hard to achieve. It is the opinion of the authors that the process should place a greater emphasis on the role of the experts. The method proposed in this paper is to sophisticate knowledge via the cycle of conversation between a deductive learning engine and a human. In this method, Genetic Programming (GP) is used as a deductive learning engine which provides great flexibility for interaction. GP shows a human the resulting learnt knowledge, of which he/she previously had only vague understanding before seeing. Next, he/she operates a fitness function of GP through choosing knowledge which he/she regards as good, to reflect them process of learning. In other words, it can be said that conditions are determined dynamically in the interaction between system and human. The authors conducted an experiment to prove the merit of this method using a prototype system.