A Study of Diversity in Multipopulation Genetic Programming

In this work we study how using multiple communicating populations instead of a single panmictic one may help in maintaining diversity during GP runs. After defining suitable genotypic and phenotypic diversity measures, we apply them to three standard test problems. The experimental results indicate that using multiple populations helps in maintaining phenotypic diversity. We hypothesize that this could be one of the reasons for the better performance observed for distributed GP with respect to panmictic GP. Finally, we trace a sort of history of the optimum individual for a set of distributed GP runs, trying to understand the dynamics that help in maintaining diversity in distributed GP.