Doctor Code: A Machine Learning-Based Approach to Program Repair

Abstract

To address the problems of automatic repair techniques, we present Doctor Code, a new APR technique that chooses repair operators by systematically learning from the features of the most common bugs in different programs, using machine learning. The wise selection of repair operators reduces the number of candidate patches. We compare our technique against Mutation repair, a test suite-based APR technique, using the Siemens suite. The experiment results indicate that our technique can fix 41 bugs while the baseline only repairs 22. In addition, Doctor Code can produce patches that do not exist in the search space of the three test suite-based techniques called SPR, Prophet, and SemFix. We also experiment with Doctor Code utilizing three buggy versions of a program called Space (9K LOC), to indicate its capability of repairing large-sized programs. In addition, we compare Doctor Code against 7 state-of-the-art APR tools like Elixir, using the Defects4j dataset. The experiment results indicate that our technique outperforms the other tools regarding the number of fixed bugs and overfitted patches.Comparing Doctor Code with RAPR as the baseline indicates that using machine learning reduces the number of overfitted patches and the time of patch production by 33.33% and 82.68%, respectively.