Paper amortizes Thompson sampling for simulator-based design
Original title: Amortized Thompson Sampling and Efficient Information-Directed Sampling for Simulator-Based Design
Precomputed approximations enable information-directed sampling and are shown on two synthetic-biology simulators with intractable likelihoods.
The paper proposes amortizing Thompson sampling before data are collected so an approximation of the sampling distribution is ready when observations arrive. The method targets multi-round sequential acquisition and also estimates expected information gain about the optimal action. Combining that estimate with immediate regret supports information-directed sampling, which the authors say can outperform Thompson sampling. They demonstrate it on two simulator-based synthetic-biology design problems whose simulators have intractable likelihoods.
Unconfirmed: Performance claims are the authors' and are not independently verified here.
- Amortization is done before any data are collected.
- The setting is multi-round acquisition and Thompson-sample recommendation.
- Expected information gain can rank samples or designs.
- Information-directed sampling balances that gain against immediate regret.
- Demos use two synthetic-biology simulators with intractable likelihoods.
From the original article
1 min read at arxiv.orgThompson sampling is a widely used strategy for sequential decision-making under uncertainty, offering a principled way to balance exploration and exploitation. This strategy involves maximizing a target objective under a single sample drawn from the current belief over unknown quantities.
By Michal Kobiela · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.