Sumie Morita

Detection of unexpected situations by applying software reliability growth models to test phases

In software development, software reliability growth models (SRGMs) often provide values that do not meet expectations; sometimes the results of the SRGM and the actual data disagree and other times the SRGM overestimates the expected values. The former often occurs in model curves and the predicted number of faults. For example, the software reliability growth curve cannot describe the situation where developers stop testing multiple times because the equations in SRGMs cannot treat such information. The latter can arise when the total number of expected faults is 100, but the SRGM indicates 1000. If developers encounter such situations, they often doubt the SRGM results and hesitate using SRGMs for predictions. In this study, we apply two different cases of SRGM. Two projects of Fujitsu Labs Ltd. are analyzed using SRGM either for the entire dataset or each test phase. Based on the results and interviews with the developers, we found that the model using separate test phases provides a better fit because faults counted in each test phase have different viewpoints and the deviation between SRGM and expectations indicates a problem with development.

Defect Analysis and Prediction by Applying the Multistage Software Reliability Growth Model

In software development, defects are inevitable. To improve reliability, software reliability growth models are useful to analyze projects. Selecting an expedient model can also help with defect predictions, but the model must be well fitted to all the original data. A particular software reliability growth model may not fit all the data well. To overcome this issue, herein we use multistage modeling to fit defect data. In the multistage model, an evaluation is used to divide the data into several parts. Each part is fitted with its own growth model, and the separate models are recombined. As a case study, projects provided by a Japanese enterprise are analyzed by both traditional software reliability growth models and the multistage model. The multistage model has a better performance for data with a poor fit using a traditional software reliability growth model.