Improved Bayesian Networks for Software Project Risk Assessment Using Dynamic Discretisation

It is possible to build useful models for software project risk assessment based on Bayesian networks. A number of such models have been published and used and they provide valuable predictions for decision-makers. However, the accuracy of the published models is limited due to the fact that they are based on crudely discretised numeric nodes. In traditional Bayesian network tools such discretisation was inevitable; modelers had to decide in advance how to split a numeric range into appropriate intervals taking account of the trade-off between model efficiency and accuracy. However, recent a recent breakthrough algorithm now makes dynamic discretisation practical. We apply this algorithm to existing software project risk models. We compare the accuracy of predictions and calculation time for models with and without dynamic discretisation nodes.