Software development in practice is still human intensive, involving human error, human judgment and subjective assessment. The outcome is often uncertain and unpredictable and, along with uncertainty in natural phenomena, inevitably leads to software risks and uncertainties. Despite the pervasiveness of human involvement, surprisingly few techniques exist that model related software uncertainties explicitly. This dissertation begins to remedy this situation by presenting the Maxim of Uncertainty in Software Engineering, stating that ``Uncertainty is inherent and inevitable in software processes and products,'' and, as corollary, urging that software uncertainties be modeled explicitly.
Ideally, an uncertainty modeling technique should include probabilistic notions of uncertainty and confidence, support multiple factors of influence, and allow dynamic updating of uncertainty values during development. Here, we claim that Bayesian Belief Networks meet these desiderata and are therefore suitable for software uncertainty modeling.
We are specifically interested in software traceability and how it may be improved through notions of judgment, confidence, belief, evidence, and causal influence. To this end, we posit a research hypothesis, stating that software uncertainty modeling, specifically Bayesian network modeling of developers' confidence levels, allows improved traceability of software artifacts and relations.
This requires that Bayesian network models of confidence levels be constructed. We describe how, given a system of interrelated software artifacts, an isomorphic Bayesian network is constructed. The structure and belief values in this network are later revised and refined as the network is validated against project data and developer judgment.
For validation, we studied the artifacts and relations of a system being developed at Beckman Instruments. Bayesian networks were constructed to capture and confirm developers' confidences in software artifacts and relationships. An example scenario of confidence fluctuation was used to study developers' ability to trace and track artifacts and associated confidences with and without the Bayesian information. The study showed that improvement in developer scores was, as hypothesized, achieved with statistical significance. Additional contributions include a prototype implementation of a Java applet for software belief networks and a collection of confidence factors that may influence levels of developer confidence in software requirements.