Abstract: All commonly used, general purpose algorithms for constructing experimental designs work design point by design point-rowwise. We introduce an algorithm that works columnwise, that is, factor by factor. In common with other algorithms, ours requires an optimality criterion with respect to a specified model. Among its advantages are its ability to accommodate a priori notions of symmetry and balance, to adapt experiments with sequentially processed factors, and to incorporate goodness-of-fit considerations. Through a series of problems, we explore the properties and utility of this approach, comparing its solutions to those of other design algorithms in terms of D-optimality, design yield, and level balance.
Key words and phrases: Balance, covariates, missing values, optimal design, semiconductor industry.