1. NEW EXPERIMENTAL DESIGNS FOR METAMODEL BUILDING.
- Author
-
Auzins, Janis and Janushevskis, Alexander
- Subjects
- *
EXPERIMENTAL design , *QUADRATIC forms , *MEAN field theory , *MATHEMATICAL optimization , *ANALYSIS of variance , *APPROXIMATION theory , *HYPERCUBE networks (Computer networks) - Abstract
This paper explores space-filling experimental designs for use with local quadratic approximations. A Mean Squared Distance (MSD) criterion is offered for optimization of the uniformity of designs. The same criterion is also called the Mean Square Error (MSE) criterion by other authors. The authors have proposed a method of optimization of designs in unit cube and spherical regions, which can be used for designs with constrained or unconstrained level values. Also, other criteria such as Entropy, Discrepancy, MaxMin and Eglaj`s criterion are compared by approximation of the test function, using several weight functions for local approximation. It is shown that the Latin Hypercube type designs, optimized according to the MSD criteria, give the best results by approximation of test function both without statistical error and with statistical error with normal distribution. The design optimization method is used also for the well-known NASA High Speed Civil Transport aircraft (HSCT) approximation challenge. For this case, the choice of a fixed number of experiments (from the given 2490) according to the MSD criteria and following local linear approximation gives good results, more accurate than with solutions proposed by other authors, using the Response surface and the Kriging methods. [ABSTRACT FROM AUTHOR]
- Published
- 2007