Agent Learning for Mixed-Initiative Knowledge Acquisition

摘要

This research has advanced the Disciple learning and problem solving theory for rapid development and maintenance of adaptable cognitive assistants in uncertain and dynamic environments. These assistants can capture, use, preserve, and transfer to other users the subject matter expertise which currently takes years to establish, is lost when experts separate from service, and is costly to replace. The research has developed an integrated set of methods for representing partially learned knowledge, for knowledge acquisition through teaching and learning, for evidence-based reasoning with Wigmorean inference networks integrating logic and probabilities, and for mixed- initiative interaction. These methods have been implemented into a general agent shell that incorporates a large amount of knowledge for evidence-based reasoning from the Science of Evidence. The agent shell can be taught to solve problems by a subject matter expert, with assistance from a knowledge engineer, to become a domain-specific cognitive assistant. The shell has been used to develop several prototype cognitive assistants. For instance, the cognitive assistant for intelligence analysis demonstrated mixed-initiative learning of complex analysis rules and mixed-initiative hypothesis analysis, allowing the analyst to drill-down based on available time and evidence, to make probabilistic assumptions, and to revise the analysis in light of new evidence.