Application of Recurrent Neural Network to Software Reliability and Defect Prediction  
Author Shadow Pritchard


Co-Author(s) Timothy Flavin; Vidhyashree Nagaraju; Lance Fiondella


Abstract Non-homogeneous Poisson process software reliability growth models assume the defect detection process can be characterized by curves based on parametric forms. However, such parametric forms may not capture all changes or details present in the data, which could limit a model's ability to predict defects accurately beyond a certain level of precision. Therefore, this paper assesses the application of simple neural network methods including artificial and recurrent neural networks to predict defects based on data collected during software testing. Two variations of recurrent networks, long short-term memory and gated recurrent unit, are considered and compared with traditional growth models to assess the predictive capability. Results suggests that RNN and its variations achieve approximately 1.6−37 and 1.8−68 times better overall goodness-of-fit measures compared to ANN and traditional NHPP models when 80% and 90% of the data is used for model fitting.


Keywords Software reliability, recurrent neural network, software failure, software reliability growth models, nonhomogeneous Poisson process, machine learning
    Article #:  RQD28-155

Proceedings of 28th ISSAT International Conference on Reliability & Quality in Design
August 3-5, 2023