Refining Fitness Functions in Test-Based Program Repair

Abstract

Genetic improvement has proved to be a successful technique in optimising various software properties, such as bug fixing, runtime improvement etc. It uses automated search to find improved program variants. Usually the evaluation of each mutated program involves running a test suite, and then calculating the fitness based on Boolean test case results. This, however, creates plateaus in the fitness landscape that are hard for search to efficiently traverse. Therefore, we propose to consider a more fine-grained fitness function that takes the output of test case assertions into account.