Preference Alignment from Choices and Response Times
summary
Combines choice and response-time data to recover latent preferences using drift-diffusion models (DDM) and log-normal race models, with a provable one-over-n convergence rate for linear DDM via stochastic gradient methods and Rademacher-complexity bounds for nonlinear drifts. Implements joint likelihood / ratio-based objectives that work under misspecification; evaluates on intertemporal choice with 100 subjects × ~140 trials (train 100 / test 40), finds lower prediction error when adding time, 95% prediction intervals cover held-out times, and estimated discount factors rise for ~90% of subjects.
AI-written summary · about
chapters (10)
- 0:00Preference alignment research
- 0:43Learning from choice data
- 2:52Choice data limitations
- 7:02Drift-diffusion models
- 18:01Biased drift-diffusion models
- 24:09Response time in learning
- 25:13Race models for choices
- 28:26Empirical evaluation
- 36:51Key takeaways
- 37:57Q&A
related talks
30mBuying Data Over Time: Approximately Optimal Strategies for Dynamic Data-Driven Decisions
18mConstantine
12mCausal Inference And Data-Driven Decision-Making
01h 08mData Driven Optimization Models and Algorithms
09mTowards Credible and Effective Data-Driven Decision-Making
44mIdentification and Estimation of Dynamic Structural Models with Unobserved Choices
55mZeroth-Order Methods for Convex-Concave Minmax Problems: Learning from Strategically Generated Data
31mInterpreting Machine Learning From the Perspective of Nonequilibrium Systems