Exploring and Exploiting Fragments

The thesis will study behaviour and uses of fragment schemas in tree-based Genetic Programming (GP). Key contributions will include: new tools for efficient analysis of the propagation of fragment schemas in GP evolution, empirically based conclusions of the validity of the GP Building Block Hypothesis (GP-BBH) with fragment schemas and the accuracy of selected results from GP schema theory, and newmethods that improve GP using fragments in genetic operators, as guides for the evolutionary process, and as program subroutines. Much research has centred around GP schema and building block theory in recent years. However, the area has become theory-rich and data-poor. Very few empirical studies have been carried out to analyze schemas in real evolutions due to the massive scale of the task. The thesis will develop a new way to approach the problem which allows empirical analysis of the fragment schemas in large populations of large programs. Using the approach will provide valuable data that will allow firm conclusions on some as yet unanswered questions in schema theory, such as whether building blocks exist in GP populations and if so, how they propagate during evolution (for instance, do they behave as predicted by the GP-BBH), and how accurate the theoretical results of schema theory are when compared to observed behaviour of schemas. Using knowledge gained from the analysis of fragment schemas in GP, the thesis will explore some of the many avenues open for improving the performance of the GP algorithm. One avenue is to use fragments as subroutines which are used in genetic programs, creating a form of Hierarchical GP (HGP) with advantages over GP with monolithic genetic programs. Improvements to the GP algorithm will be tested on a range of tasks.