Laboratory of Structural Methods of Data Analysis in Predictive
Modeling Moscow Institute of Physics and Technology
ENG
Логин:
Пароль:
Gradient-free optimization methods with ball randomization
In such applications as bi-level optimization and huge-scale optimization we can encounter a situation when we can not calculate subgradient of a convex function which is minimized. In such situations we can use stochastic approximation of the gradient using finite difference. In a recent work by Yu. Nesterov a normal randomization is considered. In this work we consider the case of randomization with vectors uniformly distributed over a unit sphere. We provide complexity bounds for simple random search for smooth convex function and strongly convex smooth convex function. Also we consider the case when the value of the function is calculated with independent stochastic error with zero mean and bounded by known value.

Авторы: Dvurechensky Pavel , Gasnikov Alexander , Анастасия Лагуновская

Дата: 27 декабря 2014

Статус: опубликована

Журнал: Abstracts of V International Conference on Optimization Methods and Applications (OPTIMA-2014)

Страницы: 59

Год: 2014

Google scholar:

Направления исследований

Structural optimization