A. S. Veprikov, E. D. Petrov, G. V. Evseev, A. N. Beznosikov
{"title":"Zero Order Algorithm for Decentralized Optimization Problems","authors":"A. S. Veprikov, E. D. Petrov, G. V. Evseev, A. N. Beznosikov","doi":"10.1134/S1064562424602336","DOIUrl":null,"url":null,"abstract":"<p>In this paper we consider a distributed optimization problem in the black-box formulation. This means that the target function <i>f</i> is decomposed into the sum of <span>\\(n\\)</span> functions <span>\\({{f}_{i}}\\)</span>, where <span>\\(n\\)</span> is the number of workers, it is assumed that each worker has access only to the zero-order noisy oracle, i.e., only to the values of <span>\\({{f}_{i}}(x)\\)</span> with added noise. In this paper, we propose a new method <span>ZO-MARINA</span> based on the state-of-the-art distributed optimization algorithm <i><span>MARINA</span></i>. In particular, the following modifications are made to solve the problem in the black-box formulation: (i) we use approximations of the gradient instead of the true value, (ii) the difference of two approximated gradients at some coordinates is used instead of the compression operator. In this paper, a theoretical convergence analysis is provided for non-convex functions and functions satisfying the PL condition. The convergence rate of the proposed algorithm is correlated with the results for the algorithm that uses the first-order oracle. The theoretical results are validated in computational experiments to find optimal hyperparameters for the Resnet-18 neural network, that is trained on the CIFAR-10 dataset and the SVM model on the LibSVM library dataset and on the Mnist-784 dataset.</p>","PeriodicalId":531,"journal":{"name":"Doklady Mathematics","volume":"110 1 supplement","pages":"S261 - S277"},"PeriodicalIF":0.5000,"publicationDate":"2025-03-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://link.springer.com/content/pdf/10.1134/S1064562424602336.pdf","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Doklady Mathematics","FirstCategoryId":"100","ListUrlMain":"https://link.springer.com/article/10.1134/S1064562424602336","RegionNum":4,"RegionCategory":"数学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q3","JCRName":"MATHEMATICS","Score":null,"Total":0}
引用次数: 0
Abstract
In this paper we consider a distributed optimization problem in the black-box formulation. This means that the target function f is decomposed into the sum of \(n\) functions \({{f}_{i}}\), where \(n\) is the number of workers, it is assumed that each worker has access only to the zero-order noisy oracle, i.e., only to the values of \({{f}_{i}}(x)\) with added noise. In this paper, we propose a new method ZO-MARINA based on the state-of-the-art distributed optimization algorithm MARINA. In particular, the following modifications are made to solve the problem in the black-box formulation: (i) we use approximations of the gradient instead of the true value, (ii) the difference of two approximated gradients at some coordinates is used instead of the compression operator. In this paper, a theoretical convergence analysis is provided for non-convex functions and functions satisfying the PL condition. The convergence rate of the proposed algorithm is correlated with the results for the algorithm that uses the first-order oracle. The theoretical results are validated in computational experiments to find optimal hyperparameters for the Resnet-18 neural network, that is trained on the CIFAR-10 dataset and the SVM model on the LibSVM library dataset and on the Mnist-784 dataset.
期刊介绍:
Doklady Mathematics is a journal of the Presidium of the Russian Academy of Sciences. It contains English translations of papers published in Doklady Akademii Nauk (Proceedings of the Russian Academy of Sciences), which was founded in 1933 and is published 36 times a year. Doklady Mathematics includes the materials from the following areas: mathematics, mathematical physics, computer science, control theory, and computers. It publishes brief scientific reports on previously unpublished significant new research in mathematics and its applications. The main contributors to the journal are Members of the RAS, Corresponding Members of the RAS, and scientists from the former Soviet Union and other foreign countries. Among the contributors are the outstanding Russian mathematicians.