A Study on the Number of Graftable Bugs in Automated Program Repair

Abstract

In recent years, automated program repair by reusing existing source code is attracting much attention. In insert operation of the automated program repair, a source code line is selected from existing source code and it is inserted to a faulty code region. However, only a limited number of bugs can be fixed by reusing code lines that exist in the source code. In this research, we propose two approaches to fix more bugs by the automated program repair. The first approach is reusing code lines from a large dataset of source code. The second approach is normalizing variable names. In this paper, we examine how much bugs can be fixed by the automated program repair. To evaluate the approaches, we examine a ratio of code lines reusable from existing source code to all inserted code lines in bug fix commits. The investigation results on 5 software repositories showed that the first and second approaches increased the ratio to 43–59% and 56–64% from 37–54%, respectively. In the case where we used both the approaches, the ratio increased to 64–69%.