Online learning, links with optimization and games
Université Paris–Saclay — 2023–2024
Thursdays 2:30–5:45 — Room 1A7
This course proposes a unified presentation of regret minimization for adversarial online learning problems, and its application to various problems such as Blackwell's approachability, optimization algorithms (GD, Nesterov, SGD, AdaGrad), variational inequalities with monotone operators (Mirror-Prox, Dual Extrapolation), fixed-point iterations (Krasnoselskii-Mann), and games. The presentation aims at being modular, so that introduced tools and techniques could easily be used to define and analyze new algorithms.
Schedule
- January, 25th
- Convexity tools
- February, 1st
- UMD Theory
- February, 15th
- Online optimization
- February, 29th
- Blackwell's approachability
- March, 7th
- Gradient methods in optimization
- March, 14th
- AdaGrad
- March, 21th
- Monotone operators and fixed point iterations
- March, 28th
- Regret learning in games
Evaluation
Evaluation will be 100% project-based.
Projects will become available progressively. A project can be attributed to at most one student. You must choose a project below (first come, first served) and notify me by sending an e-mail.
Projects have various levels of estimated difficulty, indicated by a number of stars. Grading will take that into account. If you are facing important difficulties, you can ask for help by sending me an e-mail.
Completed work must be sent by e-mail and may contain several files (e.g. a pdf and a Jupyter notebook).
The deadline is April 21st, 23:59.
Candidates for validation
- Abbadi, Omar — ★★ [theory, code] UMD-based extension of AdaGrad-Norm and application to games
- André, Pascaline — ★★ [RL, code] Adaptive diagonal scalings for Q-learning
- Benali, Nafissa — ★★ [theory, code] Solving games with line-search
- Bertrand, Virgile — ★★ [theory, code] Solving games with alternating iterations
- Biechy, Lucas — ★★★ [theory, code] AdaGrad for approachability and application to games
- Boukir, Nabil — ★★★ [theory] Alternatives to FTRL and FTL
- Bulaich Mehamdi, Abdellah — ★★ [theory] Yet another algorithm for online strongly convex optimization
- Charles, Nicolas — ★★ [theory] Composite optimization
- Gaillard, Pierre — ★★ [theory, code] Sparse payoff vectors
- Gonidec, Robin — ★ [theory] Perceptron
- Kaichouch, Elias — ★ [theory, code] A hybrid of mirror descent and dual averaging
- Papegay-Vallet, Vincent — ★★ [theory] Asynchronous fixed point iterations
- Pierron, Alex — ★★ [theory, code] Solving games: extrapolation vs optimism
- Pla, Corentin — ★★★ [theory] Variants of AdaGrad-Norm and application to games
- Prevost, Adrien — ★★ [theory] Generalized regularizers
- Reyero Lobo, Angel — ★★ [theory] Hyperbolic entropy
- Riveline, Malkiel — ★★ [theory] AdaGrad-Diagonal and generalized co-coercivity
- Roche, Matisse — ★★ [theory, code] AdaGrad on the simplex and application to games
- Saidi Alaoui, Imad — ★ [theory] Optimistic iterations with arbitrary norms for solving games
- Scarcella, Dylan — ★★★ [code] Alternative to Vovk–Azoury–Warmuth
- Schwarz, Stefanie — ★ [theory] Online linear classification with absolute loss
- Sow, Dahibou — ★ [theory] Exponential weights with step-sizes
- Taupin, Jérôme — ★★★ [theory, code] Parameter-free approachability algorithms
- Vilhes, Samy — ★★ [theory] AdaGrad-Diagonal: Stronger adaptivity to smoothness
- Weng, Weihao — ★★★ [theory] Online Newton step
- Wu, Chengyang — ★★ [theory] Regret minimization with global costs
- Xu, Minrui — ★★ [theory] Sparse bandits
- Zhang, Zhiyuan — ★ [theory] Iterates based on squared Mahalanobis norms
- Zhou, Yanyu — ★★★★ [theory, code] On the Chambolle–Pock method