Augmented D-Optimal Design for Effective Response Surface Modeling and Optimization

For effective response surface modeling during sequential approximate optimization (SAO), the normalized and the augmented D-optimality criteria are presented. The normalized D-optimality criterion uses the normalized Fisher information matrix by its diagonal terms in order to obtain a balance among the linear-order and higher-order terms. Then, it is augmented to directly include other experimental designs or the pre-sampled designs. This augmentation enables the trust region managed sequential approximate optimization to directly use the pre-sampled designs in the overlapped trust regions in constructing the new response surface models. In order to show the effectiveness of the normalized and the augmented D-optimality criteria, following two comparisons are performed. First, the information surface of the normalized D-optimal design is compared with those of the original D-optimal design. Second, a trust-region managed sequential approximate optimizer having three D-optimal designs is developed and three design problems are solved. These comparisons show that the normalized D-optimal design gives more rotatable designs than the original D-optimal design, and the augmented D-optimal design can reduce the number of analyses by 30% - 40% than the original D-optimal design.