This post describes Ader (Zhang et al. NeurIPS 2018), which achieves the optimal dynamic regret bound in OCO by running OGD with multiple learning rates and aggregating the outputs.