Mining Repair Actions for Guiding Automated Program Fixing

Automated program fixing consists of generating source code in order to fix bugs in an automated manner. Our intuition is that automated program fixing can imitate human-based program fixing. Hence, we present a method to mine repair actions from software repositories. A repair action is a small semantic modification on code such as adding a method call. We then decorate repair actions with a probability distribution also learnt from software repositories. Our probabilistic repair models enable us to mathematically reason on the automated software repair process. By applying our method on 14 repositories of Java software and 89993 versioning transactions, we show that our probabilistic repair actions are able to guide the automated fixing process in the repair space, with a probabilistic focus on likely repair shapes first