“…The adaptive construction of the neural network performs well, but it would be interesting to start with a space filling design, such as a Wootton-Sergent-Phan-Tan-Luu one (Sergent 1989), a maximin Latin Hypercube Sampling (Johnson et al 1990), or an optimal design for the promising Kullback-Leibler information criterion (Jourdan and Franco 2010), instead of the low discrepancy sequence terms. Moreover, for our chosen scenario, there are only 28 uncertain input variables and among them ten have a non-negligible impact on the IRS variability; but other configurations imply to deal with about 60 uncertain factors, and maybe up to 30 influent ones.…”