Theory Refinement Combining Analytical and Empirical Methods

Ourston, D. and R.J. Mooney, Theory refinement combining analytical and empirical methods, Artificial Intelligence 66 (1994) 273-309. This article describes a comprehensive system for automatic theory (knowledge base) refinement. The system applies to classification tasks employing a propositional Hornclause domain theory. Given an imperfect domain theory and a set of training examples, the approach uses partial and incorrect proofs to identify potentially faulty rules. For each faulty rule, subsets of examples are used to inductively generate a correction. Because the system starts with an approximate domain theory, fewer training examples are generally required to attain a given level of classification accuracy compared to a purely empirical learning system. The system has been tested in two previously explored application domains: recognizing important classes of DNA sequences and diagnosing diseased soybean plants.