Are Bayesian Networks Sensitive to Precision of Their Parameters?

In this paper, we examine whether Bayesian networks are sensitive to precision of their parameters in the context of Hepar II, a sizeable Bayesian network model for diagnosis of liver disorders. Rather than entering noise into probability distributions, which was done in prior studies, we change their precision, starting with the original values and rounding them systematically to progressively rougher scales. It appears that the diagnostic accuracy of Hepar II is very sensitive to imprecision in probabilities, if these are rounded. However, the main source of this sensitivity appears to be in rounding small probabilities to zero. When zeros introduced by rounding are replaced by very small non-zero values, imprecision resulting from rounding has minimal impact on Hepar II’s performance.