Improved Software Project Risk Assessment Using Bayesian Nets

Empirical software engineering models typically focus on predicting development effort or software quality but not both. Using Bayesian Nets (BNs) as causal models, researchers have recently attempted to build models that incorporate relationships between functionality, effort, software quality, and various process variables. This thesis analyses such models and, as part of a new validation study, identifies their strengths and weaknesses. A major weakness is their inability to incorporate prior local productivity and quality data, which limits their applicability in real software projects. The main hypothesis is that it is possible to build BN models that overcome these limitations without compromising their basic philosophy. In particular, the thesis shows we can build BNs that capture known trade-offs and can be tailored to individual company needs. The new model, called the Productivity Model, is developed by using the results of the new validation of the existing model, together with various other analyses. These include: the results of applying various statistical methods to identify relationships between a range of variables using publicly available data on software projects; analyses of other studies; expert knowledge. The new model is also calibrated using the results of an extensive questionnaire survey of experts in the area. The thesis also makes a number of other novel contributions to improved risk assessment using BNs, including. • A model which predicts the proportions of different types of defects likely to be left in software after testing. The model uses the results of statistical analysis on the past software projects. It can be combined with other defect prediction models to predict the number of residual defects of different categories. • A learning model for predicting the number of defects found and fixed in successive testing iterations.