A taxonomy of global optimization methods based on response surfaces (Q1850837)

The author presents a taxonomy of the various approaches for using response surfaces for global optimization. A distinction is made according as whether these approaches are non-interpolating or interpolating. The author demonstrates the unreliability of approaches using non-interpolating surfaces by showing that the usual methods like fitting quadric surfaces may not sufficiently capture the shape of the function. From among the basis-function methods, a distinction is made according as whether the basis-functions are fixed or the basis-functions have parameters that are tuned (kriging). The latter method enables one to make a statistical interpretation thereby allowing the user to construct an estimate of the potential error in the interpolator. The paper presents in all seven methods and brings out the importance of kriging.