Online Learning of Non-stationary Sequences

We consider an online learning scenario in which the learner can make predictions on the basis of a fixed set of experts. We derive upper and lower relative loss bounds for a class of universal learning algorithms in- volving a switching dynamics over the choice of the experts. On the basis of the performance bounds we provide the optimal a priori discretiza- tion for learning the parameter that governs the switching dynamics. We demonstrate the new algorithm in the context of wireless networks.