In this paper we present a general methodology for handling uncertain
knowledge in expert systems, which is based upon objective probability
theory. The use of objective probabilities helps to overcome some of
the difficulties in the subjective Bayesian approach. The basic idea i
s to refine a qualitative assessment of uncertainty made by a domain e
xpert into a quantitative objective probability by measuring frequenci
es in data sets. Knowledge is represented as a probabilistic network w
here the structure is elucidated from the experts, and the probability
distributions are estimated from a set of representative samples from
the domain. We test the hypothesis of independence between variables
using linear regression analysis techniques. Having identified depende
ncies we modify the structure of the network to account for them. We h
ave tested our methodology by implementing an expert system for provid
ing diagnostic advice during colon endoscopy. Our results show strong
empirical evidence supporting our approach.