A framework for memetic optimization using variable global and local surrogate models

We propose a framework of memetic optimization using variable global and local surrogate-models for optimization of expensive functions. The framework employs the trust-region approach but replaces the quadratic models with the more general RBF ones. It makes an extensive use of accuracy assessment to select the models used and to improve them if necessary. It also employs several efficient and stable numerical methods to improve its performance. Rigorous performance analysis shows the proposed framework significantly outperforms several existing surrogate-assisted evolutionary algorithms.