Coffee-House Designs

Abstract

Designs that attempt to cover the experimental region as uniformly as possible, so called space-filling designs, have regained much interest after their successful linkage to D-optimum type designs by Johnson et al. (1990). The proposed maximin distance designs, however, are very difficult to generate, especially when the region is irregular or/and high dimensional and the number of design points is large.

(Coffee-house) Designs (see Müller, 1998) which are constructed by sequentially adding maximin distance points are asymptotically equivalent and reasonably efficient. A variant with good projective properties will be proposed and compared to the approach by Morris and Mitchell (1995). A relation to a design algorithm for random coefficient regression, cf. Fedorov and Müller (1989), is revealed.