Using semantic data integration to create reliable rule-based systems with uncertainty

An expert system is considered to be reliable if it generates reliable hypotheses. The quality of the hypotheses depends mainly on the effectiveness of system's knowledge base. This paper discusses the problem of designing effective knowledge bases for rule-based systems with uncertainty. The knowledge is acquired from aggregate data stored in various repositories. The data can differ, to some extent, both in syntax and in semantics. The first part of an algorithm for rules' generation and refinement operates by means of semantic data integration. It allows to join aggregate data from different repositories and generate strong production rules. The second part of the algorithm is based on a formal concept of the normal base form. For having the property of normality, a knowledge base has to be internally consistent and not redundant. In the process of rules' refinement, the rules violating the normality are eliminated. The effectiveness of the obtained knowledge base, dependent on the base's size and on rules' reliabilities, is high. The considerations are illustrated with medical examples.