Software Reliability Growth Modeling: Comparison between Non-Linear- Regression Estimation and Maximum-Likelihood-Estimator Procedures

Automotive software complexity has been growing rapidly with time. The demand for automation in automotive segment including autonomous automobiles and software based products has caught the attention of researchers. Hence, it is necessary to check the complexity of automotive software and their reliability growth. Testing in the field of software artifact is resource intensive exercise. If project managers are able to put forward testing activities well then the testing resource consumptions may be much more resource/cost efficient. Reliability can be estimated during testing phase of software using software reliability growth models (SRGMs). A software package Computer Aided Software Reliability Estimation (CASRE) has many important SRGMs. These SRGMs are based on Non-Homogeneous Poisson Process (NHPP), Markov process or Bayesian models. Computer Aided Software Reliability Estimation-CASRE is an open source software that has been used to compare the reliability estimates using different models for a automotive software failure dataset along-with, comparison of different methods to parameter estimation (MLE and NLR). Reliability estimation can also be performed after testing phase to predict latent faults and also assess maturity of automotive software. For parameter estimation of SRGMs, two techniques are widely used, namely maximum likelihood and method of least squares. The two techniques under comparative estimation include Maximum Likelihood Estimator (MLE) and Non-Linear Regression (NLR) estimators. Assessment of prediction accuracy using relative error metric, i.e. Balanced Prediction Relative Error (BPRE), is reported lower than 5%. Further in the paper we compare these two estimation procedures for their usability and applicability in correlation with SRGMs. The data used for this study is time-domain failure. In the data, software faults have been reported with their time-between-failures (TBF). Results obtained highlight the fact that NLR is a reasonable estimator for fitting the data to observed failure data, while MLE is a better estimator for making reliable predictions.