Mixed-Initiative Exception-Based Learning for Knowledge Base Refinement

Abstract

Over the years we have developed the Disciple approach for the rapid development of knowledge bases and knowledge-based agents, by subject matter experts, with limited assistance from knowledge engineers (Tecuci et al, 2000). This approach relies on a Disciple learning agent that can be trained to solve problems by an expert. First, however, a knowledge engineer has to work with the expert to define the object ontology of Disciple. This ontology consists of hierarchical descriptions of objects and features from the application domain. Then, the expert can teach Disciple to solve problems in a way that resembles how the expert would teach a student. For instance, the expert defines a specific problem, helps the agent to understand each reasoning step toward the solution, and supervises and corrects the agent’s behavior, when it attempts to solve new problems. During such mixed-initiative interactions, the agent learns general problem solving rules from individual problem solving steps and their explanations of success or failure. A critical role in this multistrategy rule learning process is played by the object ontology, which is used as the generalization hierarchy.