Victor Picheny : May 24, 2012
Optimization using partially converged simulation : an algorithmic solution based on Gaussian Process emulation
Victor PICHENY, CERFACS
Thursday, May 24, 2:00 p.m. in the CERFACS conference room
Abstract:
In the context of expensive numerical experiments, a promising solution to alleviate the computational costs consists of using partially converged simulations instead of exact solutions. The gain in computational time is at a price of precision in the response.
Here, we propose an optimization algorithm based on Gaussian Process (GP) emulation, where the GP is used simultaneously for modeling the error due to partial convergence and for choosing the sequence of observations.
The presentation is organized as follow: first, we recall the basics of GP-based optimization; then, we describe our error model and the associated learning issues; finally, we propose a complete algorithmic solution and illustrate it on toy problems.
Here, we propose an optimization algorithm based on Gaussian Process (GP) emulation, where the GP is used simultaneously for modeling the error due to partial convergence and for choosing the sequence of observations.
The presentation is organized as follow: first, we recall the basics of GP-based optimization; then, we describe our error model and the associated learning issues; finally, we propose a complete algorithmic solution and illustrate it on toy problems.
