Online Mirror Descent
An online learning algorithm based on Bregman divergence that minimizes the instantaneous loss while keeping decisions close to the previous ones. For convex Lipschitz losses it guarantees cumulative regret upper bounds of O(√T) or O(log T).
Timeline (last 90 days)
The authors put the regularizer itself directly into the loss function of the online learner, rather than using its linearized approximation, thereby preserving convex structural information and achieving low regret.
Online mirror descent (OMD) uses the Bregman divergence as the distance measure, and for convex Lipschitz-continuous loss sequences it guarantees an upper bound on cumulative regret of O(√T) or O(log T).
All Facts (2)
The authors put the regularizer itself directly into the loss function of the online learner, rather than using its linearized approximation, thereby preserving convex structural information and achieving low regret.
50%UnverifiedOnline mirror descent (OMD) uses the Bregman divergence as the distance measure, and for convex Lipschitz-continuous loss sequences it guarantees an upper bound on cumulative regret of O(√T) or O(log T).
50%