Adaptive dynamic surrogate-assisted evolutionary computation for high-fidelity optimization in engineering

Abstract

Surrogate-assisted evolutionary computation have received much more attention in the field of optimization because of its ability to reduce effectively the CPU cost. However, as the dimension of the search space increases, the number of samples required to construct the global accurate surrogate model will increase exponentially. In order to improve the computational efficiency for achieving high-precision optimization in high-dimensional search space, an adaptive dynamic surrogate-assisted evolutionary computation approach based on variable search region is presented in this paper. The basic idea of the present method is to abandon the high accurate approximation of surrogate model on the global search space which requires a large number of samples, but focus on the search of the smaller local region where the optimal solution is located. Then refine the samples to construct a higher-precision local surrogate model on the smaller local search region, which moves with the movement of the current optimal solution. Numerical experiments show that the highly accurate optimal solution can be obtained by less than five times of adaptation. It is also applied successfully to the aerodynamic shape design optimization of transonic airfoil and wing, and the results show that present adaptive approach greatly improves the computational efficiency by about ten times compared with the traditional static global approximation surrogate model.