487. Performance Estimation framework for a range of stochastic first-order methods.
Invited abstract in session MD-8: Systematic and computer-aided analyses III: noisy gradient methods and fixed-point algorithms, stream Systematic and computer-aided analyses of optimization algorithms.
Monday, 16:30-18:30Room: B100/7007
Authors (first author is the speaker)
| 1. | Anne Rubbens
|
| ICTEAM, UCLouvain | |
| 2. | Julien Hendrickx
|
| ICTEAM |
Abstract
We rely on a computer-assisted technique to automatically obtain worst-case performance guarantees for a variety of stochastic first-order methods, and illustrate how it enables improvements over existing guarantees on several numerical examples.
This technique allows dealing with a wide range of stochastic settings, including e.g. unified variance model and finite-sum optimization, and can be applied to the analysis of non variance reduced (e.g. SDG) and variance reduced (e.g. SAGA) methods.
Keywords
- Computer-aided algorithm analysis
- Stochastic optimization
- Conic and semidefinite optimization
Status: accepted
Back to the list of papers