Elbert, Michael A.; Howe, Richard C.; Weyant, Thomas F. (1992, ASQC) Digital Equipment Corporation, Tewksbury, MA 01876
The authors examine a large and complex industrial software product to determine software reliability.
Using a tool called SMERFS (Statistical Modeling and Estimation of Reliability Functions for Software), they chose seven model to investigate for use. They then gathered data from their company's Quality Assurance Reports and Software Problem Reports and filtered it to ensure validity. They performed twenty different analyses of the seven models for each data set and calculated instantaneous defect-finding/MTBE curves for each week of project testing/month of project operation.
They discovered that the Generalized Poisson and Geometric models provided the best fit for the error count and time-between errors data. Mathematical modeling, they found, is useful for reliability measurement, although they caution that models recommended fit best in the described situation and may not be applicable to every problem. Thus you should make sure that your model matches reality and passes goodness-of-fit tests before relying on it. They warn that the customer's view of defects is worse than the producer's (number of times any given error is encountered vs. number of errors found). Performance of the reliability models depends heavily on the testing effort.
Thus mathematical modeling, if examined and used carefully, can predict software reliability.
Geometric model,Modeling,Poisson distribution,Quality assurance (QA),Reliability,Software