Comparing Genetic Programming Approaches for Non-Functional Genetic Improvement
Résumé
Genetic improvement (GI) uses automated search to find improved versions of existing software. While most GI work use genetic programming (GP) as the underlying search process, focus is usually given to the target software only. As a result, specifics of GP algorithms for GI are not well understood and rarely compared to one another. In this work, we propose a robust experimental protocol to compare different GI search processes and investigate several variants of GP-and random-based approaches. Through repeated experiments, we report a comparative analysis of these approaches, using one of the previously used GI scenarios: improvement of runtime of the MiniSAT satisfiability solver. We conclude that the test suites used have the most significant impact on the GI results. Both random and GP-based approaches are able to find improved software, even though the percentage of viable software variants is significantly smaller in the random case (14.5% vs. 80.1%). We also report that GI produces MiniSAT variants up to twice as fast as the original on sets of previously unseen instances from the same application domain.
Origine | Fichiers produits par l'(les) auteur(s) |
---|