Improving performance of automatic program repair using learned heuristics

Abstract

Automatic program repair offers the promise of significant reduction in debugging time, but still faces challenges in making the process efficient, accurate, and generalizable enough for practical application. Recent efforts such as Prophet demonstrate that machine learning can be used to develop heuristics about which patches are likely to be correct, reducing overfitting problems and improving speed of repair. SearchRepair takes a different approach to accuracy, using blocks of human-written code as patches to better constrain repairs and avoid overfitting. This project combines Prophet's learning techniques with SearchRepair's larger block size to create a method that is both fast and accurate, leading to higher-quality repairs. We propose a novel first-pass filter to substantially reduce the number of candidate patches in SearchRepair and demonstrate 85% reduction in runtime over standard SearchRepair on the IntroClass dataset.