An automatic knowledge acquisition tool

Abstract

We describe a large-scale automatic knowledge acquisition tool, the Quasi-Optimizer (QO) system. It is a domain-independent program that can obtain, verify, fuse and optimize human expertise. The QO system is capable of generating computer models (descriptive theories) of human decision making strategies in two ways. In the passive mode of observation, the system does not or cannot interfere with the environment and records the characteristic features of the situations and the corresponding strategy responses to them. In the active mode of observation, the system designs a sequence of environments (“experiments”) for the decision making strategy to respond to. The design of the experiments can be fixed in advance or can follow a dynamically evolving pattern that minimizes the total number of experiments needed for a userspecified level of precision. A module of QO can ascertain whether a non-static strategy is in fact learning, that is converging to an asymptotic form, to which the program can then extrapolate. Another module can assign a quality measure (credit) to the different strategy components identified by it, on the basis of their short- or long-term benefits in attaining certain goals. The QO system can select the best components of several strategies and combine them in a Super Strategy. The inconsistencies, incompletenesses and redundancies inherent in such a Super Strategy are eliminated and a Quasi-Optimum Strategy is generated. This strategy is better, in the statistical sense, than any one participating in the “training set”. The Quasi-Optimum Strategy, therefore, corresponds to a normative theory within the limitations of the available information. We also describe two rather different areas of application — a possible approach to the automation of air traffic controllers' training and evaluation, and the automatic verification and validation of discrete-event simulation models.