Knowledge acquisition for probabilistic expert systems

The probabilstic formalism has been used in creating the REFEREE system, a belief-network-based expert system designed to aid readers in determining the credibility of a randomized clinical trial. The effect the formalism had on the process of knowledge acquisition based on this experience is explored. Although the system is still in development, several of those effects are reported. Specifically, the need to make operational definitions of concepts deemed important to the expert forced the organization of a domain that was formulated initially for a rule-based system. Categorizing probability distributions as being logical, probabilistic, or prototypical helped to decrease the number of probability assessments. On the other hand, the lack of an intermediate prototype may have prolonged development, and computational limitations forced occasional compromises.