DISH: A Distributed Hybrid Optimization Method Leveraging System Heterogeneity

arXiv2212.02638MaRDI QIDQ6419632FDOQ6419632

Publication date: 5 December 2022

Abstract: We study the minimax optimization problems that model many centralized and distributed computing applications. Existing works mainly focus on designing and analyzing specific methods, such as the gradient descent ascent method (GDA) and its variants or Newton-type methods. In this work, we propose GRAND as a gradient-related ascent and descent algorithmic framework for solving minimax problems. It allows update directions within acute angles to the partial gradients. GRAND covers gradient-type, Newton-type, scaled gradient, and other general descent ascent methods as special cases. We also propose its alternating version, Alt-GRAND, in which

x

and

y

are updated sequentially. For distributed computing problems, GRAND enables flexible methods' designs to better exploit agents' heterogeneous computation capabilities in a system. In particular, we propose hybrid methods allowing agents to choose gradient-type or Newton-type updates based on their computation capabilities for both distributed consensus and network flow problems. For both GRAND and Alt-GRAND, we present global sublinear and linear rates in strongly-convex-nonconcave and strongly-convex-PL settings, respectively. These results match the state-of-the-art theoretical rates of GDA in corresponding settings and ensure the linear convergence of distributed hybrid methods. We further discuss the local superlinear performance of related Newton-based methods. Finally, we conduct numerical experiments on centralized minimax and distributed computing problems to demonstrate the efficacy of our methods. To the best of our knowledge, GRAND is the first generalized framework for minimax problems with provable convergence and rate guarantees. Moreover, our hybrid methods are the first to allow heterogeneous local updates for distributed computing under general network topology.

This page was built for publication: DISH: A Distributed Hybrid Optimization Method Leveraging System Heterogeneity

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6419632)