【2026-10-02】Dr. Julian Zimmert, Google Research, "Change of Measure in Online Linear Optimization"

 

Abstract:

Online Linear Optimization (OLO) is a fundamental problem solved with canonical algorithms such as Follow the Regularized Leader (FTRL) or Online Mirror Descent (OMD).

Many sequential decision problems are solved by reducing them to OLO. While this strategy has proven extremely successful, regret components external to OLO sometimes lead to suboptimal overall regret. In a surprising number of distinct problems, a change-of-measure would immediately resolve this issue.

This talk presents the linear change-of-measure technique for FTRL and OMD. Although this technique isn't new, it is surprisingly little known and authors continue to propose complicated solutions to problems that are trivially resolved using change-of-measure. 

 

Bio:

Julian Zimmert is a Senior Research Scientist at Google Research based in the AI Center of Google Berlin. He is staying at NTU as a Visiting Scholar for the 2026/2027 academic year. Julian received his Ph.D. (2020) from the Department of Computer Science at the University of Copenhagen (Denmark). He is best known for his foundational work on adversarially robust Multi-armed Bandits, which inspired a rich literature on robust algorithms in Bandits and Reinforcement Learning. At Google, he developed a principled method to cost-efficiently crawl the web in the modern ecosystem.

He is interested in Online Learning with a special focus on partial information problems such as Reinforcement Learning and Bandits.

Julian also serves as the Disability Representative of Google Berlin and is interested in how Machine Learning can help with accessibility and inclusion.