Hedging in games: Faster convergence of external and swap regrets

Authors: Xi Chen, Binghui Peng

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical For two-player games, we show that the regret of optimistic Hedge decays at rate O(1/T 5/6), improving the previous bound of O(1/T 3/4) by Syrgkanis, Agarwal, Luo and Schapire [27]. In contrast, we show that the convergence rate of vanilla Hedge is no better than O(1/ T), addressing an open question posed in Syrgkanis, Agarwal, Luo and Schapire [27]. For general m-player games, we show that the swap regret of each player decays at O(m1/2(n log n/T)3/4) when they combine optimistic Hedge with the classical external-to-internal reduction of Blum and Mansour [6]. Via standard connections, our new (swap) regret bounds imply faster convergence to coarse correlated equilibria in two-player games and to correlated equilibria in multiplayer games. Our main technical lemma behind Theorem 5.1 shows that strategies produced by the algorithm of [6] with optimistic Hedge moves very slowly in ℓ1-norm under the adversarial setting (which in turn allows us to apply a stability argument similar to [27]).
Researcher Affiliation Academia Xi Chen Department of Computer Science Columbia University xichen@cs.columbia.edu Binghui Peng Department of Computer Science Columbia University bp2601@columbia.edu
Pseudocode No The paper describes algorithms mathematically (e.g., updating rule for Hedge) but does not provide structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statements about making code open-source or include links to code repositories for the described methodology.
Open Datasets No This is a theoretical paper and does not mention the use of any datasets for training.
Dataset Splits No This is a theoretical paper and does not mention any training/validation/test splits or cross-validation.
Hardware Specification No This is a theoretical paper; no computational experiments requiring specific hardware specifications were performed or reported.
Software Dependencies No This is a theoretical paper. It does not mention any software dependencies with specific version numbers as it does not involve computational implementation details.
Experiment Setup No This is a theoretical paper. It does not describe an experimental setup with hyperparameters or system-level training settings.