Designing Efficient Computer Experiments for Metamodel Generation

, ,
,

Keywords: , , , ,

We demonstrate the relative merits of IZ-optimal designs for computer experiments relative to Latin hypercube designs, using two examples, both in two factors and eleven points, one a simple function from Welch et al. [1] and the other a micro-flowmeter response presented earlier by Crary et al. [2]. The relative reductions in variance of fitting are large, viz., 1.85 and 2.24, respectively. We show that, contrary to prevailing views, IMSE-optimal designs can have proximal design points, which we call “twin points,” with the interesting property that the design is, in part, determined by the information resources available. We comment on this newly established link between optimality and information content. We outline a particle-interaction theory and demonstrate that ground-state twin points are observed in a simple particle-dynamics simulator, when the inter-particle potential is repulsive and growing with range, and there is a centrally directed single-particle force.

PDF of paper:


Journal: TechConnect Briefs
Volume: 1, Technical Proceedings of the 2001 International Conference on Modeling and Simulation of Microsystems
Published: March 19, 2001
Pages: 132 - 135
Industry sector: Sensors, MEMS, Electronics
Topic: Modeling & Simulation of Microsystems
ISBN: 0-9708275-0-4