<ul data-eligibleForWebStory="true"><li>Non-stationary online learning has been a focus, but mainly on convex functions, neglecting stronger curved losses like squared or logistic loss.</li><li>A new approach leveraging mixability, a property to capture loss curvature for improved regret, is proposed.</li><li>Utilizing exponential-weight method with fixed-share updates, the dynamic regret is reduced to O(dT^(1/3)PT^(2/3)logT) for mixable losses.</li><li>This approach outperforms the previous best-known result, reducing the dynamic regret complexity in terms of d.</li><li>The improvement is due to an analytical framework based on mixability, which simplifies the analysis process compared to previous methods.</li></ul>

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

Discover more