Cluster evolution strategies : enhancing the sampling density using representatives

Most randomized search methods can be regarded as random sampling methods with a (non-uniform) sampling density function. Differences between the methods are reflected in different shapes of the sampling density function and in different adaptation mechanisms that update this density function based on the observed samples. We claim that this observation helps in getting a better understanding of evolutionary optimizers. An evolutionary algorithm is proposed, that uses an enhanced selection mechanism which uses not only fitness values but also considers the distribution of samples in the search-space. After a fitness based selection, the individuals are clustered, and a representative is selected for each cluster. The next generation is created using only these representatives. The set of representatives is usually small and the efficient incorporation of local search techniques is possible.