Victor Picheny: April 7, 2011

Optimization of noisy computer experiments using Gaussian processes

Victor Picheny, Laboratoire MAS, Ecole Centrale Paris

Thursday, April 7, 4:00 p.m. in the CERFACS conference room

Abstract:

We will address the issue of global optimization of expensive-to-evaluate stochastic simulators using kriging. Many of these simulators depend on factors that tune the level of precision of the response, the gain in accuracy being at a price of computational time.
First, we will recall the main principles of the state-of-the-art kriging-based algorithm for deterministic functions, EGO (Efficient Global Optimization). Then, we will show how to adapt this algorithm to the noisy case, using a quantile-based criterion that allows a rigorous treatment of heterogeneous response precisions.
Secondly, we will present a procedure that takes advantage of the flexibility of many stochastic simulators, by allocating on-line the computational time given to each measurement, allowing a better distribution of the computational effort and increased efficiency.
The proposed algorithm will be illustrated by an application in nuclear criticality safety.