Concept Generation from Relational Structure

Abstract

Discovery learning, which acquires new concepts or knowledge, is one of the most advanced forms of machine learning. Few systems have been proposed for discovery learning in practical use, and most of them are based on various heuristics. Discovery learning is considered to consist of two processes: inductive acquisition of general structure(relational structure) from existing knowledge base, and application of the relational structure to a domain knowledge for acquiring new concepts in the domain. In this paper we mainly focus on the application process, and propose a method of generating new concepts, that is, new predicates which do not occur in the domain knowledge, by applying the relational structure. We prove that the new generated clauses including the new predicates are consistent with the domain knowledge, and propose an algorithm for approximate calculating the new clauses from the relational structure and the domain owledge in finite steps. We give proof of some useful theorems for this algorithm. In addition, we discuss the the method for the acquisition of relational structure from an existing knowledge base.