A Hybrid Bayesian Network Modeling Environment

Bayesian networks are a powerful method for building probability models. But t he formalism does not support i ncremental model development and reuse of models. This is partly due to the fact that Bayesian n etworks require precise probability values, while incremental model development and model reuse require the ability to abstract probability information. We present a formalism called hybrid Bayesian n etworks that combines the traditional formalism with that of qualitative probabilistic networks (5). Qualitative probabilistic networks represent probability information with signs showing directionality of influence between random variables. Our formalism allows a model builder to start by specifying only qualitative influences and then add quantitative information as it is available and as time permits. The modeling environment can infer bounds on u nspecified probability values based on those specified and on the type of qualitative influence.