Collaborative Development of Large Bayesian Networks

Bayesian networks (BN) have been shown to be a very effective form of models for diagnostic assistants. However, the difficulties creating a BN model for complex domains have been a barrier to their use. We present a methodology and supporting software for rapid and robust development of BN models for diagnostic systems. Our approach uses a layered structure of BN and custom node types. These reduce the complexity of the models without degrading their fidelity. Together with keyword tags and policy based conflict resolution, they make it possible to develop subsystem models that are merged into a single integrated model. Extracting and re-merging the subsystem models allows cyclic development by appropriate domain experts. We have developed an editor for the model that can be used directly by the domain expert without assistance of a knowledge engineer. The expert enters the domain information into simple tables. The BN file used for reasoning and other BN computations is created automatically by the editor.