Development of Pi Sigma Neural Network Model for the Prediction of Software Reliability Using 5 NASA Failure Datasets

Abstract

Software reliability models are usually used to model the failure of software systems and prediction of its reliability potential. These models are however plagued with less accuracy, efficiency, and resource-effectiveness. Some soft computing methods have not yet been implemented to investigate their effectiveness and robustness for software fault prediction. Pi Sigma Neural Network (PSNN) software reliability prediction model was developed in this study for a better understanding of the modelling of software systems defects and reliability validated on 5 NASA promise datasets after carrying out data analysis using Seaborn on Python, working with raw data, pre-processed data with min-max normalization, Synthetic Minority Oversampling Technique (SMOTE) to overcome class imbalance problem between defective and non-defective modules, and then correlational analysis with varying thresholds (0.8, 0.85, 0.9 and 0.95) to reduce noise and get key features. The results obtained using the PSNN model showed for all the datasets good average performance for recall being highest at 79.8% based on no threshold, precision at 76.2% on 0.9 threshold, f1-score with 75.6% on 0.95 threshold and accuracy at 74.8% with the same 0.95 threshold. A model based on recall is good at fault finding. Modifying the structure and architecture of the PSNN, like using a voting ensemble algorithm of varied combinations of PSNNs and using a firefly algorithm to optimize in the future, will improve the Neural Network technique.