Empirical analysis of different levels of meta-evolution

We analyze different levels of meta-evolution using a graph based GP system. The system allows one to represent individuals of the search space and genetic variation operators in a coherent way as graph programs differing only in the operator set. Seven variants of meta-evolution are tested on three real world classification problems. The most complex variant consists of three meta-levels where graph programs on meta-level 1 recombine individuals of the search space (base level), graph programs on meta-level 2 recombine programs on meta-level 1, and programs on meta-level 3 recombine programs on meta-level 2 and themselves. The empirical results shows that the use of meta levels is advantageous.