Online Convex Optimization

Sorbonne Université - M2 ISDS (2023 - 2024)

Content of the class

In complex environments where simple stochastic models and traditional statistical methods may not be applicable, such as in scenarios like spam detection where spammers continuously challenge spam filters, a robust approach is needed. This approach involves learning from ongoing experiences as new problem aspects unfold. This is the essence of online convex optimization, where data is processed in real-time, feedback is continuously incorporated, and algorithms are dynamically updated. Online convex optimization has gained significant attention, particularly due to its relevance in internet-related applications. These applications encompass tasks like ad selection, recurrent auctions, spam detection, expert/algorithm aggregation (boosting), and more. This course aims to introduce and explore the fundamental concepts of online convex optimization and develop algorithms with rigorous theoretical analysis.
Prerequisite: probability theory (notion of random variables, convergence of random variables, conditional expectation), linear algebra.

Teachers

Pierre Gaillard and Joseph de Vilmarest.

Evaluation

Details on the evaluation criteria will be provided here as soon as possible.

Homework

Details on the homework will be provided here as soon as possible.

Reading material

The class will follow these lecture notes from O. Wintenberger, 2022. Some very related monograph is Introduction to online convex optimization by E Hazan.