Exploiting the Curvature of Feasible Sets for Faster Projection-Free Online Learning

Abstract: In this paper, we develop new efficient projection-free algorithms for Online Convex Optimization (OCO). Online Gradient Descent (OGD) is an example of a classical OCO algorithm that guarantees the optimal $O(\sqrt{T})$ regret bound. However, OGD and other projection-based OCO algorithms need to perform a Euclidean projection onto the feasible set $\mathcal{C}\subset \mathbb{R}<sup>d$ whenever their iterates step outside $\mathcal{C}$. For various sets of interests, this projection step can be computationally costly, especially when the ambient dimension is large. This has motivated the development of projection-free OCO algorithms that swap Euclidean projections for often much cheaper operations such as Linear Optimization (LO). However, state-of-the-art LO-based algorithms only achieve a suboptimal $O(T<sup>{3/4})$ regret for general OCO. In this paper, we leverage recent results in parameter-free Online Learning, and develop an OCO algorithm that makes two calls to an LO Oracle per round and achieves the near-optimal $\widetilde{O}(\sqrt{T})$ regret whenever the feasible set is strongly convex. We also present an algorithm for general convex sets that makes $\widetilde O(d)$ expected number of calls to an LO Oracle per round and guarantees a $\widetilde O(T<sup>{2/3})$ regret, improving on the previous best $O(T<sup>{3/4})$. We achieve the latter by approximating any convex set $\mathcal{C}$ by a strongly convex one, where LO can be performed using $\widetilde {O}(d)$ expected number of calls to an LO Oracle for $\mathcal{C}$.