Several reliability models are compared on the basis of their estimation capability and their dependence on particular reliability data. The relation- ship of interest is the increase in the cumulative number of discovered errors as a function of the time spent testing or using software. The author considers the exponential, hyperexponential, and various S-shaped reliability models. The estimating capabilities of the models are tested against error data sets from seven projects.
This study is a contribution to the software reliability field because it goes beyond simply stating which model performed best as an estimator. The author considers features of the user’s environment that make one model a better choice than another. Some features are refreshingly practical: What data do you have to work with? For example, is it fault detection time or failure detection time? Is it execution time or calendar time? The practical orientation of this paper will make it of interest both to researchers and users of software reliability models.