ABSTRACT
Inductive learning is one of the most powerful techniques for constructing knowledge-based systems. It enables full (or semi-)automatic formalization of the knowledge base. In other words, if we Specify proper knowledge representation, hypotheses and enough examples, i.e. proper inductive bias (Utgoff, 1986), we do not need to determine the details of the knowledge. Quinlan (1986) proposed an efficient induction algorithm named ID3 to generate classification knowledge in the form of a decision tree from examples represented in a vector form, which consists of a class to which an example belongs, the features that it has and the corresponding values that it satisfies. Although it provides quite reasonable knowledge statistically, the learned trees are not yet developed enough from the expert's point of view. That is because the system knows nothing about the target domain, so the induced knowledge sometimes lacks a lot of essential constraints required in the domain.
