Online Mirror Descent

Online Mirror Descent (OMD) is a family of online learning algorithms aiming to minimize regret, the difference between an algorithm's cumulative loss and that of the best fixed strategy in hindsight. Current research focuses on extending OMD's applicability to diverse settings, including imperfect-information games, decentralized data scenarios, and problems with delayed feedback, often incorporating techniques like optimistic updates and adaptive learning rates. These advancements improve convergence speed and robustness, impacting fields such as multi-agent systems, online advertising, and federated learning by providing efficient and theoretically sound algorithms for decision-making under uncertainty. The development of parameter-free and computationally efficient variants further broadens OMD's practical impact.