Neural Change Prediction maps code changes to behavior without execution

Researchers Laura Plein, Souhila Zidane, Jordan Samhi, and Andreas Zeller present Neural Change Prediction, a technique that learns associations between software changes and their dynamic effects by automatically applying numerous mutations and observing how output changes. The method requires executing the program under test for each mutation. From these (change, effect) pairs, models are created that (1) for a desired behavior change predict where and how to change the code, and (2) for a given code change predict its effect on output. The technique does not require prior knowledge of the code or its semantics. Case studies on CSS configuration files and Python programs demonstrate generality and wide applicability.