509. Analysis of derivative-free algorithms on noisy problems
Invited abstract in session TB-1: Zeroth-Order Optimization Methods for Stochastic and Noisy Problems, stream Zeroth and first-order optimization methods.
Tuesday, 10:30-12:30Room: B100/1001
Authors (first author is the speaker)
| 1. | Alexandre Chotard
|
| Université du Littoral de la Côte d'Opale | |
| 2. | Anne Auger
|
| Inria and Ecole Polytechnique |
Abstract
In the context of derivative-free optimization, we discuss in a first part of the talk the robustness to noise of a few optimizers that include quasi-Newton algorithms (used with finite differences to estimate the gradient), the Nelder-Mead algorithm as well as stochastic or randomized algorithms like Evolution Strategies.
In a second part of the presentation we discuss proven sufficient conditions on the step-size adaptation scheme of a simple Evolution Strategy to ensure the linear convergence on the class of scaling-invariant functions perturbed with multiplicative noise.
Keywords
- Stochastic optimization
- Optimization under uncertainty
- Derivative-free optimization
Status: accepted
Back to the list of papers