Adaptive Causal Experimental Design: Amortizing Sequential Bayesian Experimental Design for Causal Models

ICLR 2025 Conference Submission12881 Authors

Keywords: Causal Discovery; Causal Reasoning; Bayesian Experimental Design; Amortized Variational Inference; Expected Information Gain; Mutual Information Lower Bounds; Adaptive Experiments

TL;DR: A novel policy-based Bayesian sequential design method that generates adaptive, non-myopic interventions for flexible causal queries by maximizing variational Expected Information Gain lower bound, without performing intermediate Bayesian inference.

Abstract:

Interventions are essential for causal discovery and causal reasoning. Acquiring interventional data, however, is often costly, especially in real-world systems.
A careful experimental design can therefore bring substantial savings.
In the sequential experimental design setting,
most existing approaches seek the best
interventions in a greedy (myopic) manner that does not account for the synergy from the yet-to-come future experiments. We propose Adaptive Causal Experimental Design (ACED),
a novel Bayesian sequential design framework for learning a design policy capable of generating non-myopic interventions that incorporate the effect on future experiments.
In particular, ACED maximizes the Expected Information Gain (EIG) on flexible choices of causal quantities of interest (e.g., causal structure, specific causal effects) directly, bypassing the need for computing intermediate posteriors in the experimental sequence.
Leveraging a variational lower bound estimator for the EIG, ACED trains an amortized policy network that can be executed rapidly during deployment.
We present numerical results demonstrating ACED's effectiveness on synthetic datasets with both linear and nonlinear structural causal models, as well as on in-silico single-cell gene expression datasets.

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12881