|
Abstract:
|
In many scientific and engineering settings, the response surface of a physical phenomenon or engineering system has complex characteristics. The goal is to identify input settings that explore interesting features of the underlying surface. We present new sequential design methodology and are interested in developing computationally efficient methodology that detects and fits the underlying boundary curve, which usually has a complex shape. The proposed methodology utilizes a variety of techniques, such as regression trees, structural change models, various tests such as Anderson-Darling and likelihood ratio tests and QQ-plots. We start with a small initial design, which is subsequently refined, as more information about the boundary curve is obtained. The methodology proves fairly robust to noise and is illustrated on a number of simulated data sets.
|
