A framework for multi-model EDAs with model recombination

Estimation of Distribution Algorithms (EDAs) are evolutionary optimization methods that build models which estimate the distribution of promising regions in the search space. Conventional EDAs use only one single model at a time. One way to efficiently explore multiple areas of the search space is to use multiple models in parallel. In this paper, we present a general framework for both single- and multi-model EDAs. We propose the use of clustering to divide selected individuals into different groups, which are then utilized to build separate models. For the multi-model case, we introduce the concept of model recombination. This novel framework has great generality, encompassing the traditional Evolutionary Algorithm and the EDA as its extreme cases. We instantiate our framework in the form of a real-valued algorithm and apply this algorithm to some well-known benchmark functions. Numerical results show that both single- and multi-model EDAs have their own strengths and weaknesses, and that the multi-model EDA is able to prevent premature convergence