Automatic patch generation via learning from successful human patches

Abstract

Automatic patch generation holds out the promise of automatically correcting software defects without the need for developers to manually diagnose, understand, and correct these defects. This dissertation presents two novel patch generation systems, Prophet and Genesis, which learn from past successful human patches to enhance the patch generation process. The core of Prophet and Genesis is a novel learning technique that extracts universal properties of correct code and a novel inference technique that generalizes universal patching strategies across different applications. The results show that the learning and inference techniques enable Prophet and Genesis to operate with rich and tractable search spaces that contain many useful patches and efficient search algorithms that prioritize correct patches. By collectively leveraging development efforts worldwide, Prophet and Genesis automatically generate correct patches for real-world defects in large open-source C and Java applications with up to millions lines of code.