SVR Enhanced Kriging for Optimization with Noisy Evaluations

Numerical noise is an unavoidable by-product of Computational Fluid Dynamics (CFD) simulations, which bring challenges to optimizations. In the former work, we have proposed the ε-kriging model that can adaptively filter the numerical noise in the sample data by adding the insensitive factor (ε) of a support vector regression (SVR) model to the diagonal of the correlation matrix of a kriging model. Here we aim to develop the surrogate optimization method based on it for tackling the problems with noisy evaluations. The infilling criterion is developed to guide global optimization. It is compared with the classical kriging based optimization for couples of benchmark problems varying nonlinearity and dimension, with noise of low, medium and high intensity. The results show that our method successfully converged to the global optimums no matter how strong the numerical noise is. Drag minimization of NACA0012 airfoil also obtained satisfactory results. The results indicate that our method is effective and robust for optimizations affected by noise.