A unitary distributed subgradient method for multi-agent optimization with different coupling sources

TL;DR

This paper introduces a distributed subgradient method called ${\rm DSA_2}$ for multi-agent convex optimization with multiple local objectives, achieving non-ergodic convergence and extending to coupled constraints via dual decomposition.

Contribution

The paper proposes a novel distributed subgradient algorithm with non-ergodic convergence for multi-agent optimization, including extensions to coupled constraints using dual methods.

Findings

Achieves $O(1/\sqrt{t})$ convergence rate in objective error.

Ensures convergence of local minimizing sequence without ergodic averaging.

Validates theoretical results with numerical experiments.

Abstract

In this work, we first consider distributed convex constrained optimization problems where the objective function is encoded by multiple local and possibly nonsmooth objectives privately held by a group of agents, and propose a distributed subgradient method with double averaging (abbreviated as ${\rm DSA_2}$) that only requires peer-to-peer communication and local computation to solve the global problem. The algorithmic framework builds on dual methods and dynamic average consensus; the sequence of test points is formed by iteratively minimizing a local dual model of the overall objective where the coefficients, i.e., approximated subgradients of the objective, are supplied by the dynamic average consensus scheme. We theoretically show that ${\rm DSA_2}$ enjoys non-ergodic convergence properties, i.e., the local minimizing sequence itself is convergent, a distinct feature that cannot be found in existing results. Specifically, we establish a convergence rate of $O(\frac{1}{\sqrt{t}})$ in terms of objective function error. Then, extensions are made to tackle distributed optimization problems with coupled functional constraints by combining ${\rm DSA_2}$ and dual decomposition. This is made possible by Lagrangian relaxation that transforms the coupling in constraints of the primal problem into that in cost functions of the dual, thus allowing us to solve the dual problem via ${\rm DSA_2}$. Both the dual objective error and the quadratic penalty for the coupled constraint are proved to converge at a rate of $O(\frac{1}{\sqrt{t}})$, and the primal objective error asymptotically vanishes. Numerical experiments and comparisons are conducted to illustrate the advantage of the proposed algorithms and validate our theoretical findings.