A Discussion on Generality and Robustness and a Framework for Fitness Set Construction in Genetic Programming to Promote Robustness

A significant problem in Genetic Programming consists in ensuring the robustness or generality of the evolved code (its ability to work correctly on never before seen data). We examine approaches attempted so far, and propose a different solution, based on multiple training sets (to discriminate between possible interpretations), augmentation and refinement, which improves on both task specification and distribution of fitness. We then present some preliminary results on its application to a fairly canonical GP problem (wall-following behavior) that has been previously shown to be very brittle. Finally, we discuss issues in using additional information about the generality of individuals.