Towards Automated, End-to-End Software Defect Prediction  
Author Rashid Mijumbi

 
Co-Author(s) Kazuhira Okumoto; Abhaya Asthana

 
Abstract Current trends towards softwarization of telecommunication systems, coupled with efforts to use such systems to support advanced applications (such as autonomous driving), have re-emphasised the importance of software quality assurance. To ensure that high quality software is delivered on schedule, reliability growth models are typically used to predict the total number of residual defects. However, state-of-the-art approaches either require significant amounts of training data, involve many parameters that are difficult to estimate, or cannot accurately predict defects during all phases of the software development process. In this paper, we present a software defect prediction method that is both accurate, easy to model, and covers the entire software development process. The proposed method involves the sequential application of three novel models. First, the early defect prediction model (eDPM) uses a quantile plot of the feature development plan and the found defects to predict defects starting early on in the development process. Closer to software delivery, the eDPM is augmented by an automated, piece-wise, exponential software reliability growth model (SRGM) so as to produce a stable and smooth prediction. Finally, the post-delivery defect prediction model (pDPM) 􀀀 a variation of eDPM which uses data from a previous release rather than feature data 􀀀 is used to predict defects after software delivery. Using defect data from four releases of a large radio access network software project, we illustrate the need for each of the three models, as well as the accuracy of the end-to-end solution.

 
Keywords practical software reliability assurance; postdelivery defect prediction; software reliability modelling
   
    Article #:  RQD25-162
 
Proceedings of 25th ISSAT International Conference on Reliability & Quality in Design
August 1-3, 2019 - Las Vegas, NV, U.S.A.