Interval-Based Uncertainty Handling in Model-Based Prediction of System Quality

Our earlier research indicates feasibility of applying the PREDIQT method for model-based prediction of impacts of architecture design changes on system quality. The PREDIQT method develops and makes use of the so called prediction models, a central part of which are the "Dependency Views" (DVs) - weighted trees representing the relationships between system design and the quality notions. The values assigned to the DV parameters originate from domain expert judgments and measurements on the system. However fine grained, the DVs contain a certain degree of uncertainty due to lack and inaccuracy of empirical input. This paper proposes an approach to the representation, propagation and analysis of uncertainties in DVs. Such an approach is essential to facilitate model fitting, identify the kinds of architecture design changes which can be handled by the prediction models, and indicate the value of added information. Based on a set of criteria, we argue analytically and empirically, that our approach is comprehensible, sound, practically useful and better than any other approach we are aware of.