Expert System Framework for Adaptive Evidential Reasoning: Application to In-Flight Reoute Re-Planning

摘要

The successful introduction of AI technology into Air Force avionics has been hindered by the need for intelligent real-time reasoning with incomplete and often inconsistent data. The primary objective of the present research has been to explore the feasibility of new methods for inference in avionics expert systems, which address specific shortcomings in existing approaches and combine some of their distinct virtues. A variety of approaches have been critically reviewed: quantitative representations (Bayes, Shafer, and Zadeh), qualitative frameworks (e.g., Doyle, Toulmin, P. Cohen), and efforts to synthesize logic and probability (Nilsson, Sage). An innovative framework for expert system reasoning has been developed which combines quantitative manipulation of uncertainty (via Shaferian belief functions), a qualitative frame for representing an evidential argument, and a non-monotonic capability for revising probabilistic arguments when they lead to conflicting results. This framework, along with a personalized user interface, has been implemented in a small-scale demonstration system for in-flight responses to pop-up threats, the Adaptive Route Replanner (ARR). Results with ARR strongly confirm the feasibility of a system which reasons intelligently and flexibly in the face of uncertainty.