Efficient global optimization using multiple infill sampling criteria and surrogate models

Efficient global optimization (EGO) can be used to dramatically improve the efficiency of a design optimization based on high-fidelity and expensive numerical simulations such as computational fluid dynamics (CFD). In order to further improve the efficiency and take the advantage of parallel computing, this paper propose a parallel infill sampling criterion for EGO. Instead of adding a single sample point in each updating cycle like the original EGO does, the EGO with the proposed method can obtain arbitrary number of new sample points per cycle, which are to be evaluated in parallel. And different from most existing parallel infill criteria, such as kriging believer method, the proposed method employs different infill criteria simultaneously per cycle. Due to combination of various infill criteria, each criterion’s inherent drawbacks may be complemented. The proposed method is verified by a numerical example (two-dimensional Rastrigin function without constraints) and demonstrated with a constrained drag minimization of RAE2822 airfoil parameterized with 18 design variables. The results show that the optimization efficiency is significantly improved compared to the serial EGO and the effectiveness is promoted as well in contrast to the existing parallel infill criteria.