Software Reliability Prediction Based on Least Squares Estimation

Abstract

Of the main interest in practical software testing would be the prediction of the number of detected faults in future rather than the estimation of the quantitative reliability function, which is defined as the probability that the software does not fail during a pre-specified time interval after release. In other words, practitioners wish to know the number of remaining faults and the saturated level of the cumulative number of faults, in order to control the software testing process. In not only such a situation but also the case where the likelihood information of statistical models is not fully available, the least squares estimation (LSE) may be simple but very useful in estimating model parameters. Hence, the LSE methods should be comprehensively studied in the context of software reliability engineering. In this paper we examine four LSE methods with application to the software reliability prediction. In real data analyses based on several software fault data, we show that LSE methods are still attractive in terms of goodness-of-fit performance and predictive performance in many cases. Also, we analyze the software reliability models with trend change point based on the LSE methods.