For technological companies to stay competitive, they offer their customers new and stable software, which they ensure with quality control. Due to rapid development, despite robust testing, errors occur, errors that can be detected through statistical testing. To better understand the problem, the introduction contains a presentation of the discussed network software and clearly defined goals for statistical testing. Hypotheses are based on the number of alarms (count of reboots) generated at a given software version. Due to the count-like nature of alarms, the work focuses on methods of modeling count data. The focus was on the use of generalised linear models with the addition of modeling excessive zeroes. We concluded the work with an empirical analysis based on the number of restarts, where we performed an evaluation of individual models and compared their quality. Among all the models, the negative binomial model fitted best, in this case, the additional modeling excess of zeros did not yield major improvements. The model showed that for different software versions, the differences between the number of reboots were statistically significant.
|
