This paper presents an efficient evolutionary search strategy based on design of experiments, genetic algorithms and response surface modelling. The strategy is constructed around a genetic algorithm while incorporating elements from design of experiment (DoE) and Kriging. In particular, the design points used to update the approximation model are derived from two surfaces, one is the approximation itself which provides the prediction of the function and the other is based on the error surface computed from posterior error estimates of the Kriging model. A genetic algorithm, which supports clustering, is used on both surfaces to return multiple points for parallel evaluation of the true function. A screening method is also used to remove points lying close to existing points based on the correlation coefficients between the point to be evaluated and all existing points. Numerical experiments suggest that significant improvements can be achieved using the proposed approach. Applications of the approach on engineering design problems are also studied.
展开▼
