Adaptive Online Learning with Varying Norms

Abstract
Given any increasing sequence of norms , we provide an online convex optimization algorithm that outputs points in some domain in response to convex losses that guarantees regret where is a subgradient of at . Our method does not require tuning to the value of and allows for arbitrary convex . We apply this result to obtain new "full-matrix"-style regret bounds. Along the way, we provide a new examination of the full-matrix AdaGrad algorithm, suggesting a better learning rate value that improves significantly upon prior analysis. We use our new techniques to tune AdaGrad on-the-fly, realizing our improved bound in a concrete algorithm.
View on arXivComments on this paper