Journals Proceedings

Abstract

In modern engineering practice computer simulations are often used as a substitute for laboratory experiments. This setup yields a simulation-driven optimization problem in which the computer simulation acts as the objective function. Since simulation runs are typically computationally expensive, metamodels are used to approximate the simulation. However, this setup can face difficulties in high-dimensional problems as the accuracy of metamodels then becomes very poor. To address this issue this paper proposes an optimization framework which incorporates a dimensionality-reduction technique into the search. This allows to formulate a valid lower dimensional problem which is easier to solve. The solution found is then mapped back to the original high dimensional space. Performance analysis with an airfoil shape optimization problem shows the effectiveness of the proposed framework.