Buying Data Over Time: Approximately Optimal Strategies for Dynamic Data-Driven Decisions
summary
Model where a one-dimensional hidden state follows a drifting Gaussian state and noisy samples cost a per-sample fee plus optional fixed setup cost; decision-maker schedules purchases under an average budget and trades guessing (quadratic loss) versus a constant outside option. Analysis uses Kalman filtering to track posterior variance and shows bursty collection can beat uniform sampling: on-off policies are asymptotically optimal with zero fixed cost, and lazy policies guarantee at least half the optimal gain when fixed costs exist.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:18DEMAND FOR DATA
- 1:11QUESTIONS
- 1:54A DESIGN PROBLEM
- 2:27A MOTIVATING EXAMPLE
- 7:01THIS TALK
- 8:45WARMUP: A STATIC MODEL
- 12:25A DYNAMIC MODEL
- 13:52DATA COLLECTION POLICY
- 15:49ANALYSIS
- 24:37ASYMPTOTIC OPTIMALITY
- 26:40APPROXIMATE OPTIMALITY
- 28:17CONCLUDING THOUGHTS
related talks
54mDynamically Aggregating Diverse Information
34mSimple Pricing Schemes for Consumers with Evolving Values
34mPandora's Box with Correlations: Learning and Approximation
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
44mIdentification and Estimation of Dynamic Structural Models with Unobserved Choices
45mPreference Alignment from Choices and Response Times
01h 08mData Driven Optimization Models and Algorithms
55mAlgorithmic Trading and Machine Learning