Designing Adaptive Experiments For Policy Learning And Inference
summary
Adaptive experiments address multi-armed bandit problems but require pre-specified design choices: use top-two Thompson Sampling with probability floors and large initial batches, simulate feasible worst-case arm gaps to set sample size, and separate a learning then evaluation phase or use on-policy evaluation batches to get unbiased estimates. Plain sample means are negatively biased; use Horvitz-Thompson inverse-probability weighting or doubly robust estimators (or stabilized Hayek weights) for unbiasedness and valid inference, and document algorithms and parameters explicitly.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 1:55How to design adaptive experiments
- 13:47Alternative Approach
- 20:59Example
- 24:03Algorithms
related talks
48mAdaptivity and Confounding in Multi-armed Bandit Experiments
41mKeynote: Personalization, Bandits, and Causal inference
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
30mPractical Multi Armed Bandits
39mModern Bayesian Experimental Design
01h 03mExperimentation in Data Science
01h 04mIntroduction to Experimental Design
01h 15mBridging Stochastic and Adversarial Bandits