On the performance of a software reliability model for grouped failure data

In previous publications, an alternative prediction model, namely, compound Poisson (CP) software reliability model in software testing was studied. Later, two different estimation methods, i) compound Poisson maximum likelihood estimation (CPMLE), and ii) compound Poisson nonlinear regression (CPNLR) estimation methods were derived for time-failure data, generally in terms of CPU seconds. For time-between failure data, the competing Musa Okumoto (MO) method usually proved better due to little clumping within each CPU second. However, the proposed compound Poisson model proved better for grouped failure data where within each hour, day or week, the data accumulated considerably to the advantage of the compound Poisson. Owing to the non-influencing of the failures within sampling time-units (hour, day or week), the compounding distribution was popularly selected to be geometric, hence the name Poisson geometric. In addition to the ARE (averaged relative error) measure, Kolmogorov-Smirnov (K-S) test statistics are employed for comparing the goodness of fit to the actual empirical distribution. The K-S statistics indicated that for grouped failure data (weekly data sets WD1-WD5), CPMLE or CPNLR performed better.<
>