Industrial experimental design since BH$^2$
Wednesday 4 June, 9:00--10:00

Organizer: Scott VANDER WIEL
Bell Labs
Moderator: Don SUN
Bell Labs

Industrial experimental design since BH$^2$

C.F.J. WU
Department of Statistics, University of Michigan, Ann Arbor, Michigan, USA
There have been several major advances in industrial experimental design since the 1978 publication of the influential book by Box, Hunter and Hunter. The most obvious one was brought about by the work of Taguchi on robust parameter design and his style of doing practical industrial work. Because of the new(as of 1985) emphasis on variation reduction through the explicit use of noise factors in the experiment, new paradigms and methodologies have been proposed in the last decade. I will first review the major work in this area on performance measures, planning techniques, and modeling and analysis strategies for robust design experiments. Another major change is the extensive use of designs(i.e., experimental plans) that have mixed-levels and complex aliasing among their effects. The original motivation for their use was run size economy and flexibility. "Traditional wisdom" dictated that these designs be used only for screening purposes (ie, for estimating factor effects.) It turns out that they can also be used for estimating a small to moderate number of interactions. Analysis strategies and supporting design-theoretic work will be presented. Returning to the 2-level and 3-level series of regular fractional factorial designs, the minimum aberration criterion has emerged as the major criterion for selecting optimal fractional factorial designs. Some results and practical implications on the use of this criterion will be presented. The coverage of BH$^2$ on non-normal data is minimal. As the last topic I will survey the recent work on analysis of non-normal data from fractionated experiments. Typically these strategies exploit the design structure and should be distinguished from standard regression strategies that do not take the design structure into account.

[C.F.J. Wu, Department of Statistics, University of Michigan, Ann Arbor, Michigan, jeffwu@stat.lsa.umich.edu]


Back to JRC program
Back to JRC Home Page