Non-stationary Online Learning for Curved Losses: Improved Dynamic Regret via Mixability

Open in new window