Bootstrap selection of multivariate additive PLS spline models

In this paper we present a way of constructing design of experiments by Multivariate Additive Partial Least-Squares Splines, in short MAPLSS. Trying to select the most informative model based on an as small as possible training data set, we process an adaptive incremental selection of observations by a particular bootstrap procedure. Why MAPLSS models? Because they inherit the advantages of the boosted PLS regression that permits to capture additively non-linear main effects and relevant interactions in the difficult framework of small training sets. The effectiveness of this approach is illustrated on the reservoir simulator data used to forecast oil production.