Country-Scale Bandit Implementation for Targeted COVID-19 Testing
summary
Implemented a country-scale testing system called AVA that uses an optimistic Gittins-style index with pseudo-Gittins indices (certainty-equivalent updates) and Bayesian Beta priors to allocate 6–8k daily tests among 30–100k arrivals, handling 48-hour lab delays via a two-week rolling horizon. Port-specific greedy assignment and LASSO identify risky cities beyond country-level. Off-policy evaluation shows about 80% more positives than random testing; policy actions include graylisting 10 countries after detected spikes.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:46Policy Levers
- 2:35Goals of EVA
- 3:20The Eva System
- 5:14Exploration-Exploitation Tradeoff
- 6:26Optimistic Gittins Index
- 7:48Challenges
- 10:36Batched Bandits
- 11:41Certainty-Equivalent Updates
- 12:35Within a Batch
- 13:57Allocations in a Batch
- 16:12Testing Constraints
- 17:17Algorithm
- 20:50Empirical Bayes
- 23:24Arm Granularity
- 24:59Feature Selection
- 26:13Preliminary Estimates of Benefits
- 28:15Other Benefits
- 29:03Unreliable Public Data
- 30:26Thank you!
related talks
23mWhat cyber security can teach us about COVID-19 testing
01h 01mEfficient and Targeted COVID-19 Border Testing via Reinforcement Learning
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
41mKeynote: Personalization, Bandits, and Causal inference
17mByron Galbraith, Chief Data Scientist, Talla, NYC 2017
30mPractical Multi Armed Bandits
27mBeyond a Machine Learning Model
37mAI/ML Strategies & Techniques for a Post Pandemic World