Improved crossover operators for genetic programming for program repair

Abstract

GenProg is a stochastic method based on genetic programming that presents promising results in automatic software repair via patch evolution. GenProg’s crossover operates on a patch representation composed of high-granularity edits that indivisibly comprise an edit operation, a faulty location, and a fix statement used in replacement or insertions. Recombination of such high-level minimal units limits the technique’s ability to effectively traverse and recombine the repair search spaces. In this work, we propose a reformulation of program repair operators such that they explicitly traverse three subspaces that underlie the search problem: Operator, Fault Space and Fix Space. We leverage this reformulation in the form of new crossover operators that faithfully respect this subspace division, improving search performance. Our experiments on 43 programs validate our insight, and show that the Unif1Space without memorization performed best, improving the fix rate by 34%.