Most machine learning does not exploit prior knowledge. Theory refinement (a.k.a. theory revision or knowledge-base refinement) is the task of modifying an initial imperfect knowledge-base (KB) to make it consistent with empirical data. The goal is to improve the performance of learning by exploiting prior knowledge, and to acquire knowledge which is more comprehensible and related to existing concepts in the domain. Another motivation is to automate the process of knowledge refinement in the development of expert systems and other knowledge-based systems.

We have developed systems for refining knowledge in various forms including:

• Propositional logical rules: Systems: EITHER, NEITHER
• First order logical rules (logic programs): Systems: FORTE
• Certainty-factor rules: Systems: RAPTURE
• Bayesian networks: Systems: BANNER

These systems have demonstrated an ability to revise real knowledge bases and improve learning in several real-world domains including:

• DNA promoter sequence identification
• DNA splice-junction sequence identification
• Medical diagnosis of bacterial infections
• Soybean disease diagnosis
• Student modeling for intelligent tutoring