Distributed Mirror Descent for Online Composite Optimization

Abstract

In this paper, we consider an online distributed composite optimization problem over a time-varying multi-agent network that consists of multiple interacting nodes, where the objective function of each node consists of two parts: a loss function that changes over time and a regularization function. This problem naturally arises in many real-world applications ranging from wireless sensor networks to signal processing. We propose a class of online distributed optimization algorithms that are based on approximate mirror descent, which utilize the Bregman divergence as distance-measuring function that includes the Euclidean distances as a special case. We consider two standard information feedback models when designing the algorithms, that is, full-information feedback and bandit feedback. For the full-information feedback model, the first algorithm attains an average regularized regret of order O(1/ √ T ) with the total number of rounds T . The second algorithm, which only requires the information of the values of the loss function at two predicted points instead of the gradient information, achieves the same average regularized regret as that of the first algorithm. Simulation results of a distributed online regularized linear regression problem are provided to illustrate the performance of the proposed algorithms.Index Terms-Online distributed optimization, composite objective, average regularized regret, approximate mirror descent, bandit feedback.

Achieving Efficient Distributed Machine Learning Using a Novel Non-Linear Class of Aggregation Functions

“…Assumption 1. The network G = (V, E t ) and the weight matrix P (t) satisfy the following (Yuan et al, 2020…”

“…Functions f i,t for 1 ≤ i ≤ m and t ≥ 0 are convex. Additionally, they are G l − Lipschitz (Yuan et al, 2020) . That is…”

Improving Dynamic Regret in Distributed Online Mirror Descent Using Primal and Dual Information

“…In the centralized setting, it has been established that various optimization methods including online gradient descent (Zinkevich, 2003), online dual averaging (Xiao, 2010), online mirror descent (Duchi et al, 2010), and many others (Shalev-Shwartz, 2012;Hazan, 2019) achieve an upper bound of O( √ T ) and O(log T ) on the static regret, for convex and strongly convex loss functions, respectively. The static regret of distributed online convex optimization algorithms have also been extensively studied in the literature (Hosseini et al, 2013;Mateos-Núnez and Cortés, 2014;Akbari et al, 2015;Tsianos and Rabbat, 2016;Lee et al, 2016;Yuan et al, 2020), where the same regret rates have been derived under similar convexity assumptions. However, it is not previously known whether similar results hold for the more useful dynamic regret.…”

A stochastic mirror-descent algorithm for solving AXB=C over an multi-agent system

“…As a result, distributed algorithms have attracted much research attention. Particularly, distributed optimization, which agents over the network cooperately seeks a global optimal solution, has become more and more popular [3,13,14,16,17].…”