586 – Recent Developments in Designs of Experiments and Responses Surface Models
Optimal Designs for Gamut Models
William D. Heavlin
Google Inc.
Experiments exploring directions of steepest ascent (DSAs) are isomorphic to one-factor-at-a-time (1AAT) experiments—they try runs along the DSA at varying distances, typically with intermediate factor levels (IFLs). A natural extension of the 1AAT DSA approach likewise encourages IFLs, incorporates the DSA, and includes also the factor x DSA interactions. When combined with locally weighted scatterplot smoothing (“lowess�), this defines a class of varying-coefficient models called gamut models. Our main optimal design criterion is average (“integrated�) mean square prediction error (I-optimality), which is relatively friendly to IFLs. We adapt this criterion to the lowess-based varying-coefficient gamut models and assess the resulting designs.