Adaptivity and Confounding in Multi-armed Bandit Experiments
summary
Proposes deconfounded Thompson sampling, a simple tweak that samples arms by their posterior chance of being best across the whole population (not just the current context), uses contextual features and inverse‑propensity style weights, and handles delays and week‑long nonstationarity. Shows worst‑case instances where a nonadaptive RCT is essentially optimal, where deconfounded UCB fails, and where deconfounded Thompson sampling matches RCT robustness yet attains asymptotic optimality under IID contexts. Algorithms and proofs are empirical Bayes/linear‑Gaussian friendly.
AI-written summary · about
chapters (13)
- 0:00Intro
- 1:03Interactive learning
- 2:37Efficiency benefits of adaptivity
- 5:16A core tension
- 6:38This work
- 8:46Teasers for the theorists
- 12:43Objectives in adaptive experiments
- 15:03Interpolating between objectives
- 21:54Model: decision goal and prior knowledge
- 23:36Model: information gathering
- 33:10Ex: where it gets hard
- 34:16Example: day-of-week effects
- 39:20Robustness / Efficiency
related talks
41mKeynote: Personalization, Bandits, and Causal inference
45mContextual Multi-Arm Bandit and its applications to digital experiments
32mDesigning Adaptive Experiments For Policy Learning And Inference
01h 09mStochastic Bandits: Foundations and Current Perspectives
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
01h 01mStochastic Bandits: Foundations and Current Perspectives
30mPractical Multi Armed Bandits
17mByron Galbraith, Chief Data Scientist, Talla, NYC 2017