Abstract
Software reliability growth models (SRGM) assist in software release decisions by quantifying metrics based on failure data collected during testing. Complexity of SRGM can increase significantly with increasing complexity of software, thus requiring faster and stable algorithms to identify model parameters. While previous studies have attempted to address this issue through application of machine learning (ML) algorithms, lack of sufficient data to train ML models. Moreover, it is important to assess software in an online manner as data becomes available, which is difficult since limited data is available towards the beginning of testing. Incremental learning [12] provides the possibility of applying machine learning algorithms in an online manner, however lack of large data towards the beginning of testing limits the efficiency. Therefore, this paper proposes an adaptive incremental learning that utilizes a model trained on historical data which can forecast failures for the present data. Historical data is selected based on the Granger causality test. Our results indicate that the adaptive incremental learning approach achieves significantly better accuracy on smaller sample sizes compared to simple application of neural networks.