Simplifying Parametrization of Bayesian Networks in Prediction of System Quality

Bayesian networks (BNs) are a powerful means for modelling dependencies and predicting impacts of architecture design changes on system quality. The extremely demanding parametrization of BNs is however the main obstacle for their practical application, in spite of the extensive tool support. We have promising experiences from using a tree-structured notation, that we call dependency views (DVs), for prediction of impacts of architecture design changes on system quality. Compared to BNs, DVs are far less demanding to parametrize and create. DVs have shown to be sufficiently expressive, comprehensible and feasible. Their weakness is however limited analytical power. Once created, BNs are more adaptable to changes, and more easily refined than DVs. In this paper we argue that DVs are fully compatible with BNs, in spite of different estimation approaches and concepts. A transformation from a DV to a BN preserves traceability and results in a complete BN. By defining a transformation from DVs to BNs, we have enabled reliable parametrization of BNs with significantly reduced effort, and can now exploit the strengths of both the DV and the BN approach.