Researchers have resolved the minimax regret rate for the Experts Problem and Online Convex Optimization (OCO). For the d-expert problem, the minimax alternating regret is shown to be $\Theta(\log d)$, independent of the time horizon T. This marks a significant improvement over previous bounds, such as the $\mathcal{O}(T^{1/3}\log^{2/3} d)$ established by Hait et al. For general OCO over a d-dimensional compact convex set, the minimax alternating regret is $\Theta(d\log(1+T/d))$, also a substantial improvement over prior results. These findings are achieved through novel corrected variants of existing algorithms like Hedge, with carefully designed correction terms to manage unfavorable curvature. AI
IMPACT Establishes new theoretical bounds for online learning algorithms, potentially influencing future research in game theory and reinforcement learning.
RANK_REASON Two academic papers published on arXiv presenting novel theoretical results in online learning and game theory.
- Cevher
- COLT 2025
- Cutkosky
- Hait
- Luo
- Neural Information Processing Systems 2023
- OCO
- Piliouras
- Skoulakis
- Viano
- Zhang
- Cevher et al.
- Experts Problem
- Hait et al.
- Hedge
- Minimax Alternating Regret
- Online Convex Optimization
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
