Modularity and position independence in EDA-GP

There has been growing interest in Estimation of Distribution Algorithms (EDA). Conventional EDA mainly use a linear string representation, resembling an individual of Genetic Algorithms (GA). Because of the flexibility of GP style tree encoding, a limited number of researchers have started addressing estimation of distribution of GP-style tree form solutions. For simplicity, we refer to this kind of research as EDA-GP, As in conventional EDA, the focus of EDA-GP at this stage has to be finding an appropriate model. In (Shan et al., 2004), we proposed a number of criteria for an appropriate model for EDA-GP. While our focus is on EDA-GP, we note that these criteria are important not only for EDA-GP research, but may provide clues for general problem solving with tree form solutions. In this research, we empirically examine two criteria, namely modularity and position dependence. In this research, we empirically confirm their importance. Furthermore, we also validate that PRODIGY (Shan et al., 2004), the framework we propose for EDA-GP, is capable of handling it