82415

Автор(ы): 

Автор(ов): 

8

Параметры публикации

Тип публикации: 

Статья в журнале/сборнике

Название: 

Median Clipping for Zeroth-order Non-Smooth Optimization and Multi-Armed Bandit

Электронная публикация: 

Да

ISBN/ISSN: 

2331-8422

DOI: 

10.48550/arXiv.2402.02461

Наименование источника: 

  • arXiv.org

Обозначение и номер тома: 

2402.02461v5

Город: 

  • Cornell

Издательство: 

  • Cornell University

Год издания: 

2025

Страницы: 

1-30 http://arxiv.org/abs/2402.02461v5
Аннотация
In this paper, we consider non-smooth convex optimization with a zeroth-order oracle corrupted by symmetric stochastic noise. Unlike the existing high-probability results requiring the noise to have bounded κ-th moment with κ ∈ (1, 2], our results allow even heavier noise with any κ > 0, e.g., the noise distribution can have unbounded expectation. Our convergence rates match the best-known ones for the case of the bounded variance, namely, to achieve function accuracy ε our methods with Lipschitz oracle require ˜O(d2ε−2) iterations for any κ > 0. We build the median gradient estimate with bounded second moment as the mini-batched median of the sampled gradient differences. We apply this technique to the stochastic multi-armed bandit problem with heavy-tailed distribution of rewards and achieve ˜O(√dT) regret. We demonstrate the performance of our zeroth-order and MAB algorithms for various κ ∈ (0, 2] on synthetic and real-world data. Our methods do not lose to SOTA approaches and dramatically outperform them for κ ≤ 1.

Библиографическая ссылка: 

Корнилов Н.М., Дорн Ю.В., Лобанов А.В., Кутузов Н.В., Шибаев И.А., Горбунов Э.А., Назин А.В., Гасников А.В. Median Clipping for Zeroth-order Non-Smooth Optimization and Multi-Armed Bandit / arXiv.org. Cornell: Cornell University, 2025. 2402.02461v5. С. 1-30 http://arxiv.org/abs/2402.02461v5.