Resumen de Generating and Assessing Exact G-Optimal Designs

Myrta Rodríguez, Bradley Jones, Connie M. Borror, Douglas C. Montgomery

• A new algorithm is proposed for generating near G-optimal designs for second order models over cuboidal regions. Brent’s minimization algorithm is used with coordinate exchange to construct designs for two to five factors. A new graphical tool, the variance ratio fraction of design space (VRFDS) plot, is also proposed for comparison of the prediction variance of competing designs over a given region of interest. VRFDS plot comparisons between G-optimal designs and I-optimal designs show that G-optimal designs have a higher prediction variance, suggesting that for many response surface studies, G-optimal designs may not be the best choice.