[Seeding Methods for Run Transferable Libraries] Capturing Domain Relevant Functionality through Schematic Manipulation for Genetic Programming

This paper applies a recently developed technique of expression structure analysis and parametric distribution to the generation of functional content relevant to the problem domain. This functional basis set will then be iteratively sampled by a GP system as part of the Run Transferable Libraries process. We introduce a new algorithm for adapting the schematic templates discovered by such an analysis into a family of related functional expressions, differentiated by the number of arguments to the abstraction and the ordering of those arguments within the body of the function. Furthermore we investigate techniques to reduce redundancy within the family of functions generated by the parameter assignment, which consequently decreases the set of functions to be sampled by RTL, further improving performance. We validate that this technique is able to discover germane functionality within the context of the problem domain and that the library generated by this approach is competitive with a library hand designed with the optimal function set.