Projection-free Distributed Online Learning with Strongly Convex Losses

Yuanyu Wan,Guanghui Wang,Lijun Zhang

2021-01-01

Abstract: To efficiently solve distributed online learning problems with complicated constraints, previous studies have proposed several distributed projection-free algorithms. The state-of-the-art one achieves the $O({T}^{3/4})$ regret bound with $O(\sqrt{T})$ communication complexity. In this paper, we further exploit the strong convexity of loss functions to improve the regret bound and communication complexity. Specifically, we first propose a distributed projection-free algorithm for strongly convex loss functions, which enjoys a better regret bound of $O(T^{2/3}\log T)$ with smaller communication complexity of $O(T^{1/3})$. Furthermore, we demonstrate that the regret of distributed online algorithms with $C$ communication rounds has a lower bound of $\Omega(T/C)$, even when the loss functions are strongly convex. This lower bound implies that the $O(T^{1/3})$ communication complexity of our algorithm is nearly optimal for obtaining the $O(T^{2/3}\log T)$ regret bound up to polylogarithmic factors. Finally, we extend our algorithm into the bandit setting and obtain similar theoretical guarantees.