Reproduce “Stability and Robustness via Regularization: Bandit Inference via Regularized Stochastic Mirror Descent” 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
Budhaditya Halder, Ishan Sengupta, Koustav Chowdhury, Koulik Khamaru
Regularized-EXP3, a stochastic mirror descent algorithm from Rutgers University, simultaneously enables valid statistical inference via asymptotic normality, achieves near-optimal regret, and maintains robustness to o(T^(1/2)) adversarial corruptions in multi-armed bandits.
Key insights
- Algorithmic instability, which hinders valid inference in adaptive MABs, is not inherent to adaptive sampling but can be mitigated through principled regularization within an SMD framework.
- A general stability theory for SMD allows for systematic design and analysis of bandit algorithms that guarantee asymptotically valid statistical inference.
- It is possible to achieve low regret, valid statistical inference (via stability), and robustness to a non-trivial budget of adversarial corruptions simultaneously, a capability lacking in prior stable inference methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.