The ANR CHORUS project organizes a day of discussion on Statistical learning methods.
It will take place on:
May, 28th 2014, Paris.
Location: Université Paris Descartes, 45 rue des Saints-Pères, Paris 6ème – Amphi Fourier au 5ème étage University paris V
Slides will be in English but the discussions will be held in French
Presentation of the workshop
This workshop is organized within the framework of the ANR project CHORUS (“Common Horizon of Open Research on Uncertainty in Simulations”) whose objective is the development of methods and open-source softwares for uncertainty management in numerical simulations.
This workshop is a tutorial on statistical learning methods for parameter estimation and prediction with computer experiments. Most of simulation codes involve parameters that have to be estimated thanks to reference dataset and variables that model the uncertainties. Once a simulation code is “calibrated”, it can be used to predict some stochastic behavior. This workshop aims at providing theoretical and practical aspects illustrated by industrial applications.
This event will bring together the different partners of the ANR project CHORUS but is open to a wider audience.
9:00 : Welcome of the participants
9:15 – 9:30 : Introduction – G. Obozinski and N. Rachdi
9:30 – 10:45 : Tutorial on statistical learning methods, G. Obozinski (ENPC)
10:45 – 11:00 : Coffee Break
11:00 – 12:15 : Statistical learning and computer experiments, N. Rachdi (Airbus Group Innovations)
12:15-13:45: Lunch break
13:45 – 14:45 “Gaussian Processes and Sequential optimization”, N. Vayatis
14:45 – 15:30: “Delayed strains prediction in a containment building”, G. Blatman (EDF)
French nuclear reactors are enclosed by a containment building made of reinforced concrete. It is designed, in any emergency, to contain the escape of radioactive elements to a high pressure. Thus the assessment of the long-term behaviour of the concrete structure for lifetime containment has to be justified. In this purpose, a simulation strategy has been implemented, based upon a specific material constitutive law. The model is aimed at predicting the delayed strains in the containement building and depends on several input parameters.
The presentation is focused on the calibration of the model parameters with respect to experimental creep test results, using a Bayesian approach. In order to study the feasibility of the methodology, the calibration has been carried out for a simplified analytical model that was fast to evaluate. However, in order to anticipate the future use of a more realistic and complex finite element element model, the strategy has been coupled to a metamodelling scheme, namely a polynomial chaos approximation. Some results are given in terms of posterior distributions of the input parameters and of posterior confidence intervals of the simulated delayed strains.
15:30 – 15:45: Coffee Break
15:45 – 16:30: “Robust optimization for oil reservoir management”, V. Gervais (IFPEN)
To determine the optimal development planning of a reservoir (e.g. new wells position), the uncertainty on the physical properties of this reservoir need to be accounted for. To that purpose, a robust optimization method adapted to expensive simulators is used. It consists in estimating some summary statistic of the future oil production from Gaussian process surrogate models refined iteratively according to EI-based criteria. The workflow will be illustrated on a realistic reservoir case.
16:30 – 17:00: Discussion
17:00 : End of the workshop
Registration is free but appreciated.
Guillaume Obozinski & Nabil Rachdi