Fully First-Order Methods for Decentralized Bilevel Optimization

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wang, Xiaoyu, Chen, Xuxing, Ma, Shiqian, Zhang, Tong
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917181355196416
author Wang, Xiaoyu
Chen, Xuxing
Ma, Shiqian
Zhang, Tong
author_facet Wang, Xiaoyu
Chen, Xuxing
Ma, Shiqian
Zhang, Tong
contents This paper focuses on decentralized stochastic bilevel optimization (DSBO) where agents only communicate with their neighbors. We propose Decentralized Stochastic Gradient Descent and Ascent with Gradient Tracking (DSGDA-GT), a novel algorithm that only requires first-order oracles that are much cheaper than second-order oracles widely adopted in existing works. We further provide a finite-time convergence analysis showing that for $n$ agents collaboratively solving the DSBO problem, the sample complexity of finding an $ε$-stationary point in our algorithm is $\mathcal{O}(n^{-1}ε^{-7})$, which matches the currently best-known results of the single-agent counterpart with linear speedup. The numerical experiments demonstrate both the communication and training efficiency of our algorithm.
format Preprint
id arxiv_https___arxiv_org_abs_2410_19319
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Fully First-Order Methods for Decentralized Bilevel Optimization
Wang, Xiaoyu
Chen, Xuxing
Ma, Shiqian
Zhang, Tong
Optimization and Control
Machine Learning
90C06, 90C15, 90C47
This paper focuses on decentralized stochastic bilevel optimization (DSBO) where agents only communicate with their neighbors. We propose Decentralized Stochastic Gradient Descent and Ascent with Gradient Tracking (DSGDA-GT), a novel algorithm that only requires first-order oracles that are much cheaper than second-order oracles widely adopted in existing works. We further provide a finite-time convergence analysis showing that for $n$ agents collaboratively solving the DSBO problem, the sample complexity of finding an $ε$-stationary point in our algorithm is $\mathcal{O}(n^{-1}ε^{-7})$, which matches the currently best-known results of the single-agent counterpart with linear speedup. The numerical experiments demonstrate both the communication and training efficiency of our algorithm.
title Fully First-Order Methods for Decentralized Bilevel Optimization
topic Optimization and Control
Machine Learning
90C06, 90C15, 90C47
url https://arxiv.org/abs/2410.19319