Using Evidential Reasoning to Make Qualified Predictions of Software Quality

Software quality is commonly characterised in a top-down manner. High-level notions such as quality are decomposed into hierarchies of sub-factors, ranging from abstract notions such as maintainability and reliability to lower-level notions such as test coverage or team-size. Assessments of abstract factors are derived from relevant sources of information about their respective lower-level sub-factors, by surveying sources such as metrics data and inspection reports. This can be difficult because (1) evidence might not be available, (2) interpretations of the data with respect to certain quality factors may be subject to doubt and intuition, and (3) there is no straightforward means of blending hierarchies of heterogeneous data into a single coherent and quantitative prediction of quality. This paper shows how Evidential Reasoning (ER) - a mathematical technique for reasoning about uncertainty and evidence - can address this problem. It enables the quality assessment to proceed in a bottom-up manner, by the provision of low-level assessments that make any uncertainty explicit, and automatically propagating these up to higher-level 'belief-functions' that accurately summarise the developer's opinion and make explicit any doubt or ignorance.