A Dynamic Approach to Statistical Debugging: Building Program Specific Models with Neural Networks

Abstract

Computer software is constantly increasing in complexity; this requires more developer time, effort, and knowledge in order to correct bugs inevitably occurring in software production. Eventually, increases in complexity and size will make manually correcting programmatic errors impractical. Thus, there is a need for automated software-debugging tools that can reduce the time and effort required by the developer. The performance of previously developed debugging techniques can be greatly improved by combining them with machine-learning. Our research focuses on the application of neural networks within the domain of statistical debugging. Specifically, we develop methods to mine statistical debugging data that can then be used to train neural networks; these generated multi-layered neural networks can then be used to identify suspicious programmatic entities. Our developed networks are generated on a per program basis in order to leverage specific programmatic properties. In our empirical evaluation we compare our proposed approach with a state-of-the-art automated debugging technique. The results of the evaluation indicate that, for the cases considered, our approach is more effective than the considered technique.