Reproduce “Variational Bayesian Optimal Experimental Design” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Adam Foster, Martin Jankowiak, Eli Bingham, Paul Horsfall, Yee Whye Teh, Tom Rainforth, Noah Goodman
The authors develop reusable approximations for estimating expected information gain (EIG) in Bayesian optimal experimental design, sharing information across outcomes rather than separately estimating each nested quantity. In a mixed-effects benchmark, their estimator has lower bias and variance than LFIRE under a 60-second budget. In simulated preference experiments, adaptive designs reduce uncertainty faster than random designs; this result is limited to simulated agents.
Key insights
- Amortized variational inference can significantly accelerate EIG estimation by amortizing the cost of inference across different observations, leading to faster convergence.
- Variational estimators can achieve an O(T^-1/2) convergence rate to their respective (potentially biased) approximations, outperforming the O(T^-1/3) rate of NMC.
- The Variational Nested Monte Carlo (μ̂_VNMC) combines fast variational training with asymptotic consistency, allowing for initial rapid improvement followed by bias reduction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.