Poster: Program Repair that Learns from Mistakes

Automated program repair is a very active research field, with promising results so far. Several program repair techniques follow a Generate-and-Validate work-scheme: programs are iteratively sampled from within a predefined repair search space, and then checked for correctness to see if they constitute a repair. In this poster, we propose an enhanced work-scheme, called Generate-Validate-AnalyzeErr, in which whenever a program is found to be incorrect, the error trace that is the evidence of the bug is further analyzed to obtain a search hint. This hint improves the sampling process of programs in the future. The effectiveness of this work-scheme is illustrated in a novel technique for program repair, where search hints are generated in a process we call error generalization. The goal of error generalization is to remove from the search space all programs that exhibit the same erroneous behavior. The aim of this poster is to present our vision of the future of program repair, and trigger research in directions that have not been explored so far. We believe that many existing techniques can benefit from our new work-scheme, by focusing attention on what can be learned from failed repair attempts. We hope this poster inspires others and gives rise to further work on this subject.