Composite Design Criteria : Model and Variance

복합실험기준의 설정: 모형과 분산구조

  • 김영일 (중앙대학교 정보시스템학과)
  • Published : 2000.09.01

Abstract

Box and Draper( 19(5) listed some properties of a design that should be considered in design selection. But it is impossible that one design criterion from optimal experimental design theory reflects many potential objectives of an experiment, because the theory was originally based on the underlying model and its strict assumption about the error structure. Therefore, when it is neces::;ary to implement multi-objective experimental design. it is common practice to balance out the several optimal design criteria so that each design criterion involved benefits in terms of its relative "high" efficiency. In this study, we proposed several composite design criteria taking the case of heteroscedastic model. WVhen the heteroscedasticity is present in the model. the well known equivalence theorem between 1)- and C-optimality no longer exists and furthermore their design characteristics are sometimes drastically different. We introduced three different design criteria for this purpose: constrained design, combined design, and minimax design criteria. While the first two methods do reflect the prior belief of experimenter, the last one does not take it into account. which is sometimes desirable. Also we extended this method to the case when there are uncertainties concerning the error structure in the model. A simple algorithm and concluslOn follow.On follow.

원래 최적실험의 이론은 주어진 모형과 그에 따른 가정에 기초하여 발달되었기 때문에 하나의 최적실험기준이 실험이 가족 있는 여러 목적을 모두 반영하는 것이 무리이다. 따라서 실험자가 다목적 실험기준의 필요성을 느끼는 경우에는 종종 여러 최적실험 기준들의 균형을 이루는 방법을 통해 이러한 문제가 다루어진다. 본 연구에서는 이 분산 구조를 가지고 있는 모형을 예를 들어 복합적인 실험기준들을 알아본다. 왜냐하면 이분산인 경우 D-최적과 G-최적실험간의 동격이론은 더 이상 성립되지 않음에 따라 두 실험기준의 특징은 현격하게 구분되어지기 때문이다. 제약조건최적실험, 결합최적실험, 그리고 minimax 설험방법을 통한 실험기준들간의 균형을 꾀하여 보았다. 처음 두 방법은 실험자의 주관이 반영되어 실제적으로 매우 세심한 주의가 필요한 반면, minimax는 그러한 점을 해소하였다고 본다. 또한 이를 확장하여 오차의 이분산 구조에 대한 불확실성이 존재할 때 적용될수 있는 두 가지 실험기준도 마련하여 보았다. 간단한 알고리즘과 결어를 첨부하였다.

Keywords

References

  1. Statistics v.26 "Discussion" on Cook and Fedorov Atkinson, A. C.
  2. J. of the American Statistical Association v.90 D-optimum designs for heteroscedatic linear models Atkinson, A. C.;Cook, R. D.
  3. Annals of Math. Statistics v.40 Optimal and efficient design of experiments Atwood, C. L.
  4. J. of the American Statistical Association v.54 A basis for the selection of a response surface design Box, G. E. P.;Draper, N. R.
  5. Biometrika v.62 Robust Designs Box, G. E. P.;Draper, N. R.
  6. J. of the American Statistical Association v.89 On the equivalence of constrained and compound optimal designs Cook, R. D.;Wong, W. K.
  7. Statistics v.26 Constrained optimization of experimental design Cook, R. D.;Fedorov, V. V.
  8. Theory of Optimal Experiments Fedorov, V. V.
  9. Canadian J. of math v.12 The equivalence of two extreme problems Kiefer, J.;Wolfowitz, J.
  10. Error-robust statistical experimental design with application to model based sampling in auditing Kim, Y. I.
  11. Mathematische Operations Forshung und Statistik v.5 Experimental design in a class of models Lauter, E.
  12. Communications in Statistics, Part A-theory and Methods v.16 Constrained optimal designs for regression models Lee, C. M. S.
  13. J. of Quality Technology v.29 Optimal response surface designs in the presence of dispersion effects Mays, D. P.;Easter, S. M
  14. Kybernetika v.19 On hybrid experimental design Mikulecka, J.
  15. J. of the American Statistical Association v.66 Optimal experimental design for polynomial regression Stigler, S. M.
  16. J. of the Royal Statistical Society, ser. B. v.55 Heteroscedastic G-optimal designs Wong, W. K.;Cook, R. D.