[Statlist] Sémin. stat 18.02.2016 Clément Chevalier

[DIACON Corine]

SEMINAIRE DE STATISTIQUE

Institut de Statistique, Université de Neuchâtel, 2000 Neuchâtel - http://www2.unine.ch/statistics         

Jeudi 18 février 2016,  11H00,  salle C14, 1er étage, Bâtiment G (bâtiment de chimie), Faculté des sciences, Av. de Bellevaux 51 

Clément Chevalier, 
UniNE, Institut de statistique

Computer experiments and sequential sampling strategies relying on Gaussian process models

Abstract

Gaussian process (GP) models are today widely used to set-up sequential evaluation strategies of expensive computer codes in the case where the evaluation budget is drastically limited. 
For example, if the shape of a car front-rear bumper depends on, say, 3 scalar parameters, one might be interested in the values of the 3 parameters which optimize a performance in car crash test simulation. Since crashing a car (real physical experiment) or simulating the crash numerically is expensive and/or very computer intensive, a reasonable optimum needs to be obtained in very few trials.

Mathematically, if the function at hand is called f, a typical question is thus to build a greedy evaluation strategy of f which aims at finding its maximum (optimization problem). Other popular questions exist, like finding the input region where f exceeds a given threshold T (inverse problem).
In this talk, we will present a very popular GP-based sequential evaluation strategy of functions for optimization problems - which is based on the so-called Expected Improvement. We will also talk about parallel strategies, and may - in the discussions - find some common ground with survey theory. 
If we have time, we will detail sequential strategies for inverse problems.