Search for: All records

We present a nonapproximate computational method for generating ‐optimal designs in response surface methodology (RSM) settings usingGloptipoly, a global polynomial optimizer. Traditional approaches use a grid approximation for computing a candidate design's ‐score.Gloptipolycan find the global optimum of high‐order polynomials thus making it suitable for computing a design's ‐score, that is, its maximum scaled prediction variance, which, for second‐order models, is a quartic polynomial function of the experimental factors. We demonstrate the efficacy and performance of our method through comprehensive application to well‐published examples, and illustrate, for the first time, its application to generating ‐optimal designs supporting models of order greater than 2. This work represents the first non‐approximate computational approach to solving the ‐optimal design problem. This advancement opens new possibilities for finding ‐optimal designs beyond second‐order RSM models. 

Abstract The assumption of normality is usually tied to the design and analysis of an experimental study. However, when dealing with lifetime testing and censoring at fixed time intervals, we can no longer assume that the outcomes will be normally distributed. This generally requires the use of optimal design techniques to construct the test plan for specific distribution of interest. Optimal designs in this situation depend on the parameters of the distribution, which are generally unknown a priori. A Bayesian approach can be used by placing a prior distribution on the parameters, thereby leading to an appropriate selection of experimental design. This, along with the model and number of predictors, can be used to derive the D‐optimal design for an allowed number of experimental runs. This paper explores using this Bayesian approach on various lifetime regression models to select appropriate D‐optimal designs in regular and irregular design regions.