Improving software reliability prediction through multi-criteria based dynamic model selection and combination

摘要

Software reliability growth models (SRGMs) predict the software testing process by forecasting when a failure rate would fall below an acceptable threshold. However, there are certain limitations of SRGM in practice. Some of the limitations are Practitioners do not have reliable ways of determining the model in advance, Models show different performances according to adopted criteria, Relative priority of models in each criterion may change as testing proceeds, and The goodness-of-fit (GOF) performance to the observed data does not directly link to the prediction performance. The present research work has tried to present a systematic approach for making more accurate and robust software reliability prediction by decision tree learning, and a tool to support the approach. The approach involves searching the informative criteria for the prediction of the specified range. Then weights are assigned to the software reliability models according to the likelihood of better prediction and dynamically identify the models that are more likely to make the best prediction. To improve the predictive capability of the approach, the identified models are combined according to their tendencies of over- and under-prediction. The proposed approach has been evaluated using an experimental design using three steps and the results were discussed. (53 refs.)