Distributed Systems Group

This technical report introduces an extension for Estimation of Distribution Algorithms (EDAs). EDAs are evolutionary optimization methods that try to build models which estimate the distribution of promising regions in the search space. Traditional EDAs use only one single model at a time. However, a single (univariate) model can only represent a single area of the search space. After the algorithm has decided for one region of the search space, the probability that the algorithm can leave this area is very small and explore different parts of the search space are thus rarely investigated. Such an EDA will not sample new individuals in the outside from the learned model. One way to explore multiple areas of the search space is to use multiple models. Our proposed multi-model EDA uses clustering to group similar individuals. For each such group, one model is created. Additional, we introduce a model recombination operator which uses these models to create additional ones. The main idea of our approach is to prevent premature convergence by using multiple models. We suppose that, if multiple areas are explored at the same time, the chance t