A recurrent ANFIS tuned by modified differential evolution for efficient prediction of software reliability

Software failure prediction is a crucial task in software quality assurance. Although time series forecasting techniques and conventional software reliability models are used for the prediction of software reliability, very often these models fail to provide accurate predictions. Therefore, an accurate model for software reliability prediction is imperative. The objective of this paper is to introduce a hybrid model that merges a recurrent adaptive neuro-fuzzy inference system (RANFIS) and modified differential evolution (MDE) for software reliability prediction. The model employs inner spatial feedback loops and delayed output feedback to improve the conventional neuro-fuzzy system's prediction abilities while dealing with time series software failure data. The hybrid model is trained by MDE, where a new scheme of crossover and mutation has been proposed. We conduct an extensive simulation on a few publicly available benchmark datasets for computing the predicting ability of our hybrid model. Simulation results along with statistical analysis illustrate that our hybrid model predicts more precisely the time between successive failures in software and outperforms other traditional models.