Fitness Landscape Analysis applied to functional Genetic Improvement

Abstract

Genetic Improvement is the concept of a computer improving human-written code. This improves either the functional or the non-functional properties of the program. Genetic Improvement uses mutations to nd improved versions of the original program. This makes the search space for Genetic Improvement very large. Furthermore for functional improvement, the tness landscape forms large plateaus. In this thesis, we will attempt to analyse the search space of Genetic Im- provement using Fitness Landscape Analysis techniques to achieve a better understanding of the search space. To achieve this, we have edited the PyGGI framework to perform a random walk, and to analyse how large plateaus are. The PyGGI framework has been edited in such a way that it suits our needs and has such a performance that the experiments can be concluded in a reasonable amount of time. We perform the Genetic Improvement process on programs selected from the Bears benchmark, which contains many programs with bugs and test suites. The results of this thesis conclude that while the plateaus are near-innitely big, a random walk over the plateau often nds the global optimum. The only cases where the global optimum could not be found are the experiments which could not be improved with the used set of mutations. These results are in line with similar results in researches in this area.