Improving Automated Program Repair using Two-layer Tree-based Neural Networks

Abstract

We present DLFix, a two-layer tree-based model learning bug-fixing code changes and their surrounding code context to improve Automated Program Repair (APR). The first layer learns the surrounding code context of a fix and uses it as weights for the second layer that is used to learn the bug-fixing code transformation. Our empirical results on Defect4J show that DLFix can fix 30 bugs and its results are comparable and complementary to the best performing pattern-based APR tools. Furthermore, DLFix can fix 2.5 times more bugs than the best performing deep learning baseline.