Seminar: Leveraging Dual Channel Constraints to Speed Program Repair

Abstract

    Code mixes natural language in identifier names, comments, and stylistic choices (ordering and typesetting) with a formal language that defines a computation. The snippets in each language form a communication channel. Developers read both channels; a CPU processes only the formal channel. These two channels interact and constrain each other. The theory of dual channel constraints elucidates these interactions and points to their exploitation. I will illustrate how to leverage dual channel constraints to tackle the central challenge of automated program repair: finding correct patches in the vast search space of candidate patches. Real-world programs define large namespaces of variables that considerably contributes to the search space explosion. Existing program repair approaches neglect information about the program namespace, which makes them inefficient and increases the chance of test-overfitting. I will present RETE, a new program repair technique, that introduces a dual channel representation to learn project-independent information about a program’s namespace and uses it to navigate the search space of patches. On the ManyBugs and BugsInPy datasets, our experiments demonstrate that RETE generates six new correct patches that fix bugs that previous tools did not repair, an improvement of 31% and 59% over the existing state of the art.