355. Derivative free optimization with structured random directions
Invited abstract in session MC-1: Strategies to Improve Zeroth-Order Optimization Methods, stream Zeroth and first-order optimization methods.
Monday, 14:00-16:00Room: B100/1001
Authors (first author is the speaker)
| 1. | Silvia Villa
|
| Department of Mathematics, MaLGa, università di Genova | |
| 2. | Marco Rando
|
| Malga - DIBRIS, Universita degli Studi di Genova | |
| 3. | Cheik Traoré
|
| University of Genoa | |
| 4. | Cesare Molinari
|
| Università di Genova | |
| 5. | Lorenzo Rosasco
|
| DIBRIS, Universita' di Genova |
Abstract
We consider the problem of minimizing an objective function in a black-box setting where only function evaluations are available. Finite-difference methods is a class of algorithms that mimic gradient-based methods by replacing gradients with approximations built using function evaluations along a set of directions. In this talk I will focus on the specific choice of the set of directions where the directions are randomly chosen but satisfy an orthogonality constraint. I will then analyze finite-difference methods which incorporate variance-reduction techniques to minimize a finite sum of functions. I will derive convergence rates for smooth non-convex functions and show that our algorithm achieves a convergence rate that matches
the state-of-the-art. Finally,I will illustrate our method through numerical experiments.
Keywords
- Black-box optimization
- Derivative-free optimization
Status: accepted
Back to the list of papers