Research map / Probabilistic and scientific modeling / Causal discovery
Targeted Sequential Indirect Experiment Design
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an adaptive sequential design of indirect (confounded, nonlinear, multivariate) experiments that narrows kernel-based IV bounds on a targeted causal-effect query, with closed-form RKHS bound estimators.
Problem
Designing experiments that maximally inform a targeted scientific query (e.g., causal effect of a treatment) when experiments only indirectly perturb confounded, multivariate, nonlinear mechanisms, so the query is generally unidentified.
Approach
Treat experiments as instrumental variables and maintain upper/lower bounds on a linear functional of the mechanism via a minimax kernel IV objective with closed-form RKHS estimators; then adaptively update a Gaussian mixture experiment policy by REINFORCE-style gradient descent to shrink the bound gap.
Key results
- In the 2D synthetic setting with T=16 rounds, the adaptive strategy narrows bounds to essentially identify the target query Q[f0]=20 with low variance across 50 seeds.
- Explore-then-exploit outperforms the random baseline when using the same exploration distribution.
- Adaptive, EE, and AEE all improve over random; the random baseline improves only mildly over rounds.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use explore-then-exploit instead of random exploration, focusing exploitation on samples nearest the query point | Indirect experiment design · Synthetic low-dimensional setting | bound gap | Helped |
| Use alternating explore-exploit policy updates instead of random exploration | Indirect experiment design · Synthetic low-dimensional setting | bound gap | Helped |
| Adapt the experimental policy using intermediate estimates of the query bounds | Indirect experiment design · Synthetic low-dimensional setting | bound gap | Helped |
| Adapt the experimental policy using intermediate estimates of the query bounds | Indirect experiment design · Synthetic low-dimensional setting, 16 experiments | identification of target query | Helped |
| Use explore-then-exploit rather than alternating explore-exploit | Indirect experiment design · Synthetic low-dimensional setting | bound gap | No clear effect |
Limitations
- Not validated in a real-world experimental setup.
- Bounds are not provably valid for all queries due to regularization and non-convex optimization.
- Empirical validation limited to linear local functionals, Gaussian mixture policies, and synthetic settings.
Experimental setup
Datasets: synthetic data from a confounded nonlinear 2D structural model
Benchmarks: synthetic confounded nonlinear indirect experiment setting
Metrics: upper and lower bound on causal effect bound gap wallclock runtime
Baselines: random explore then exploit alternating explore exploit
Builds on: minimax instrumental variable estimation kernel IV (maximum moment restriction) REINFORCE
Closest papers by meaning
- Adaptive Instrument Design for Indirect Experiments (ICLR 2024 · similarity 0.85)
- Identifying biological perturbation targets through causal differential networks (ICML 2025 · similarity 0.85)
- Optimal Experiments for Partial Causal Effect Identification (NeurIPS 2026 · similarity 0.84)
- Fast Proxy Experiment Design for Causal Effect Identification (NeurIPS 2024 · similarity 0.83)
- When Selection Meets Intervention: Additional Complexities in Causal Discovery (ICLR 2025 · similarity 0.83)
- Smoke and Mirrors in Causal Downstream Tasks (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.