Architectural Overlay: Modifying an Architecture to Help Cognitive Models Understand and Explain Themselves

摘要

Two challenges confront efforts to develop effective intelligent systems to work in complex domains: making them easier, faster and less expensive to develop, and making them easier to comprehend, use, and trust (Haynes, Councill, & Ritter, 2004). This applies to cognitive models as well (Ritter, 2004). We have created a high-level behavior representation language and integrated development environment (IDE) called Herbal (Cohen, Ritter, & Haynes, 2005) that answers these challenges. Herbal provides explanations, it works, has been used by 60 people as the basis of a course on modeling and as a tutorial at the BRIMS conference, and the data so far suggest a 3 to 7 x performance increase compared to developing Soar code directly. The next steps are to evaluate and extend Herbal, and to apply Herbal to a real- world problem.