Quasi-Random Balance Designs for Sensitivity Analysis
Résumé
We present some advanced techniques which can be realized with Random Balance Design (RBD) for the computation of variance based sensitivity indices enhancing the precision of the first order effect calculation.The classical RBD method suffers from two classes of problems: 1) the computed values are biased with respect to the analytical values, 2) the sample design in use is not necessarily exhausting the sample space (i.e., it is unclear if the design is space filling). A freedom of choice in the design of a RBD input sample is given by the choice of the permutations. The standard RBD algorithm uses random permutations. We investigate if a clever choice of these permutation leads to a space-filling design and pays off in terms of numerical precision.Here, the permutations used to create the realizations of the input parameters are constructed from multi-dimensional quasi-Monte-Carlo (QMC) sequences. Those are then called quasi-random permutations.