Reproduce “Nearly Minimax Optimal Regret for Adversarial Linear Bandits with Strong Corruptions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Qiwei Di, Quanquan Gu

CRW-FTRL addresses finite-action linear bandits under strong corruption, where an attacker can alter feedback after seeing the sampled action and clean reward. It uses per-action radii to bound each action estimate’s sensitivity to corruption and updates its sampling distribution with follow-the-regularized-leader. The authors prove high-probability realized regret of O(√(dT log K)+dC), assuming the learner knows C; compact-action guarantees rely on inefficient discretization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.