Fuzzy Edit Sequences in Genetic Improvement

Abstract

Genetic improvement uses automated search to find improved versions of existing software. Edit sequences have been proposed as a very convenient way to represent code modifications, focusing on the changes themselves rather than duplicating the entire program. However, edits are usually defined in terms of practical operations rather than in terms of semantic changes; indeed, crossover and other edit sequence mutations usually never guarantee semantic preservation. We propose several changes to usual edit sequences, specifically augmenting edits with content data and using fuzzy matching, in an attempt to improve semantic preservation.