Recently, two general approaches have emerged in machine learning research. Empirical learning involves inductively acquiring concept descriptions by examining the similarities and differences among a large number of examples. Explanation-based learning (EBL) involves using existing knowledge to explain and generalize single examples and thereby acquire operational concept descriptions. However, both of these approaches have important weaknesses. Standard empirical methods cannot take adequate advantage of existing knowledge. Standard explanation- based methods cannot deal with domain theories which are incomplete or incorrect. Our goal is to develop robust and efficient methods that combine explanation-based and empirical techniques to refine imperfect concepts and domain theories to account for a set of empirical data.
