Sequential design of computer experiments for parameter estimation with application to numerical dosimetry

In this paper, we propose a sequential sampling approach for estimating a parameter of interest of the distribution of Y = f(X), where X has a known distribution in ℝ^d and f is an unknown, expensive-to-evaluate real-valued function. We shall adopt a Bayesian point of view which consists in modeling f as a sample of a well-chosen Gaussian process. Our global approach aims at estimating the parameter of interest with as few evaluations of f as possible. We compare our methodology with standard approaches through numerical experiments and eventually test it on real data corresponding to the exposure of a Japanese pregnant-woman model and her 26-week-old fetus to a plane wave.