Methods for Evolving Robust Programs

Many evolutionary computation search spaces require fitness assessment through the sampling of and generalization over a large set of possible cases as input. Such spaces seem particularly apropos to Genetic Programming, which notionally searches for computer algorithms and functions. Most existing research in this area uses ad-hoc approaches to the sampling task, guided more by intuition than understanding. In this initial investigation, we compare six approaches to sampling large training case sets in the context of genetic programming representations. These approaches include fixed and random samples, and adaptive methods such as coevolution or fitness sharing. Our results suggest that certain domain features may lead to the preference of one approach to generalization over others. In particular, coevolution methods are strongly domain-dependent. We conclude the paper with suggestions for further investigations to shed more light onto how one might adjust fitness assessment to make various methods more effective.