Nicolas Garland will defend his thesis
on Thursday 22thoctober 2020 at 2:00 pm
at École des Mines of Saint-Etienne, 514 room
Espace Fauriel, 29 rue Ponchardier
42100 Saint-Étienne
Céline Helbert, reporter
François Bachoc, reporter
Alain Rakotomamonjy, jury president
Rodolphe Le Riche, thesis director
Nicolas Durrande, EMSE supervisor
Jean-Antoine Désidéri
Yann Richet, invited member
Abstract
In this thesis, we aim to
improve the algorithms used by safety experts to analyse the outputs of their
computer simulations of physical phenomena.
Since simulations are expensive, the related algorithms build the design of
experiments sequentially, using metamodels to predict the outcome of the
simulations. We pay special attention to chain-produced simulations (where outputs of one
code are used as inputs in other numerical simulators).
The information given by the first codes allows a so-called
"introspective" analysis of the final quantity.
Firstly, we review different possible introspective metamodels, which treat the
results as a multi-outputs function. We thus study the different forms of co-kriging and propose an improvement for
the optimization of its hyperparameters. We also develop another introspective metamodel, the hyper-kriging. We then make a comparison between the predictive performance of these
metamodels.
Secondly, we work on the algorithms. An optimization algorithm, called "Step or Stop Optimization", is
developed, taking into account the specificities of the introspective case. It allows to stop the computation if the first steps are disengaging. Comparisons with currently used algorithm tend to confirm a significant saving
in computing resources thanks to a better consideration of the intermediate
physics of the simulation.
The strategy used in this algorithm can be generalized, and extended to be used
beyond the context of optimization problems.
