Assessing the impact of modeling limits on intelligent systems

The authors discuss validating the knowledge bases underlying intelligent systems. In particular, the concern is with identifying and remediating modeling limits that are likely to lead to unacceptable consequences in terms of system performance or safety. A general conceptual framework for considering the roles in intelligent systems of models of physical, behavioral, and operational phenomena is given. Ten types of modeling limits are identified, and their likely consequences discussed. A methodology for identifying limits in particular intelligent systems is described and illustrated via an evaluation of the pilot-vehicle interface within the Pilot's Associate. The essence of this methodology involves interactive probing and tracing of concept graph representations of knowledge bases.< >