Our research addresses the problem of developing knowledge-based agents that incorporate the knowledge of subject matter experts. Our approach is to develop a learning and problem solving agent, which can be directly taught by a subject matter expert by explaining it how to solve specific problems, and by critiquing its attempts to solve new problems. Because the accuracy of the agent's reasoning depends on the rules from its knowledge base, the process of rule improvement is very important. This dissertation presents an integrated set of methods to assist a subject matter expert in refining the rules from an agent's knowledge base, to incorporate his problem solving expertise. This dissertation presents methods to discover incompletely refined rules and to propose suggestions for their improvement; to guide the expert during the rule refinement process, focusing his attention on the reasoning steps that need to be analyzed; to refine the applicability condition of over-generalized and over-specialized rules; to modify a learned rule using a lazy refinement method; and to extend the agent's ontology to eliminate the rules' exceptions. These methods complement each other and create an integrated approach to the rule refinement problem in an evolving representation space, resulting in refined problem solving rules, which will assure a higher degree of correctness of the solutions generated by the agent. These rule refinement methods have been implemented in the Disciple learning agent shell, and have been evaluated during several experiments in complex application domains.
展开▼
