Dynamic Anisotropic Smoothing for Noisy Derivative-Free Optimization

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Reifenstein, Sam, Leleu, Timothee, Yamamoto, Yoshihisa
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916233668984832
author Reifenstein, Sam
Leleu, Timothee
Yamamoto, Yoshihisa
author_facet Reifenstein, Sam
Leleu, Timothee
Yamamoto, Yoshihisa
contents We propose a novel algorithm that extends the methods of ball smoothing and Gaussian smoothing for noisy derivative-free optimization by accounting for the heterogeneous curvature of the objective function. The algorithm dynamically adapts the shape of the smoothing kernel to approximate the Hessian of the objective function around a local optimum. This approach significantly reduces the error in estimating the gradient from noisy evaluations through sampling. We demonstrate the efficacy of our method through numerical experiments on artificial problems. Additionally, we show improved performance when tuning NP-hard combinatorial optimization solvers compared to existing state-of-the-art heuristic derivative-free and Bayesian optimization methods.
format Preprint
id arxiv_https___arxiv_org_abs_2405_01731
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Dynamic Anisotropic Smoothing for Noisy Derivative-Free Optimization
Reifenstein, Sam
Leleu, Timothee
Yamamoto, Yoshihisa
Machine Learning
Optimization and Control
We propose a novel algorithm that extends the methods of ball smoothing and Gaussian smoothing for noisy derivative-free optimization by accounting for the heterogeneous curvature of the objective function. The algorithm dynamically adapts the shape of the smoothing kernel to approximate the Hessian of the objective function around a local optimum. This approach significantly reduces the error in estimating the gradient from noisy evaluations through sampling. We demonstrate the efficacy of our method through numerical experiments on artificial problems. Additionally, we show improved performance when tuning NP-hard combinatorial optimization solvers compared to existing state-of-the-art heuristic derivative-free and Bayesian optimization methods.
title Dynamic Anisotropic Smoothing for Noisy Derivative-Free Optimization
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2405.01731