Apprenticeship Learning Techniques for Knowledge Based Systems

The significance of machine learning for the future use of computers is very great. Autonomous computer systems of the future will need far more knowledge than humans can explicitly transfer; this requires that computers learn independently. An important research goal for machine learning is to identify techniques that will allow intelligent knowledge-based systems to learn automatically the large amounts of domain-specific knowledge that are necessary for achieving expert-level problem-solving performance. This thesis describes apprenticeship learning techniques for automation of the transfer of expertise. Apprenticeship learning is a form of learning by watching, in which learning occurs as a byproduct of building explanations of human problem-solving actions. An apprenticeship is the most powerful method that human experts use to refine and debug their expertise in knowledge-intensive domains such as medicine; this motivates giving such capabilities to an expert system. The major accomplishment in this thesis is showing how an explicit representation of the strategy knowledge to solve a general problem class, such as diagnosis, can provide a basis for learning the knowledge that is specific to a particular domain, such as medicine. The Odysseus learning program provides the first demonstration of using the same technique to transfer of expertise to and from an expert system knowledge base. When watching an expert, it improves a knowledge base for the pre-existing Heracles expert system shell. When watching a student apprentice, it models the student against the knowledge base and thereby identifies bugs and gaps in the student's fledgling expertise. Another major focus of this thesis is limitations of apprenticeship learning. It is shown that extant techniques for reasoning under uncertainty for expert systems lead to a sociopathic knowledge base. A knowledge base is sociopathic if there exists a subset of the knowledge base that gives better performance than the original knowledge base. Incremental learning techniques, which subsume apprenticeship learning, have inherent limitations if a knowledge base is sociopathic. Also, the synthetic agent method is presented, which provides a means of determining a performance upper bound for an apprenticeship learning system.