Orateur
Emmanuel Gobet
(CMAP, École polytechnique, IPP)
Description
Abstract: Regression-based methods constitute a standard approach to solve dynamic programming stochastic equations. Their theoretical accuracies can be quantified in terms of local approximation errors, statistical errors and propagation errors. There is an subtle interplay between these three sources of error, which should lead to determine well the approximation space according to the sampling effort. In this talk, I will discuss
- the pros/cons of using Discontinuous Galerkin type space and Neural Network approximation spaces;
- statistical learning results to adaptively choose the approximation space.
