This paper reviews the use of Bayesian networks (BNs) in predicting software reliability and software defects. The approach allows analysts to incorporate causal process factors as well as combine qualitative and quantitative measures, hence overcoming some of the well-known limitations of traditional software metrics methods. The approach has been used and reported on by organizations such as Motorola, Siemens, and Philips. However, one of the impediments to more widespread use of BNs for this type of application was that, traditionally, BN tools and algorithms suffered from an obvious ‘Achilles’ heel’ – they were not able to handle continuous nodes properly, if at all. This forced modelers to have to predefine discretization intervals in advance and resulted in inaccurate predictions where the range, for example, of defect counts was large. Fortunately, recent advances in BN algorithms now make it possible to perform inference in BNs with continuous nodes, without the need to pre-specify discretization levels. Using such ‘dynamic discretization’ algorithms results in significantly improved accuracy for reliability and defects prediction type models.
Keywords: Bayesian Networks, Defects, Predicting, Reliability, Software
