Abstract
Many software reliability models exist, making it especially difficult for individuals new to software reliability to apply models effectively. Researchers have suggested a wide variety of statistical model selection techniques as well as frameworks that combine multiple selection techniques. However, these techniques often require a level of expertise and subjectivity that continue to make it difficult to apply software reliability models in practice. To overcome these limitations of past research, this paper presents a neural network approach to rank models based on their ability to accurately predict future defect discovery. Unbiased simulation techniques are used to generate training data as well as additional test data not used for training. Our results indicate that the model that predicts future defects most accurately was recommended by the neural network over 60% of the time and the first or second most accurate model was recommended by the neural network over 95% of the time. Moreover, in over 50% of cases where the second most accurate model was recommended, the predictive sum of squares error was no more than a factor of two times greater than the most accurate model, suggesting that neural networks may be a viable model selection approach for software reliability growth models.