Robust experiment design via stochastic approximation (Q1076462)

The authors consider the optimum design problem for estimation of parameters in a statistical model. The parameters are assumed to have a known prior distribution and a design is sought which maximizes the expected value of the determinant of the Fisher information matrix with respect to this prior distribution. Such a design is called an ED- optional design. Four stochastic approximation algorithms are proposed for iterative computation of the designs. Two of these algorithms are applied to various examples and it is found that the tuning constant in one of the algorithms is highly problem dependent but that in the second algorithm is practically problem independent in the cases tested.