<ul><li>The study revisits the Follow the Regularized Leader (FTRL) framework for Online Convex Optimization (OCO) to improve dynamic regret guarantees in dynamic environments.</li><li>The research demonstrates that with a new approach focusing on optimistic composition of future costs and linearization of past costs, FTRL can achieve known dynamic regret bounds by pruning some of the costs.</li><li>The analysis reveals that the success in optimizing dynamic regret with FTRL lies in its ability to balance between greedy and agile updates through careful cost management and minimal recursive regularization.</li><li>The findings suggest that the key to addressing dynamic regret efficiently is not in changing the fundamental nature of FTRL but in synchronizing the algorithm's state with its iterates through strategic pruning.</li></ul>

On the Dynamic Regret of Following the Regularized Leader: Optimism with History Pruning

Discover more