Research map / Reinforcement learning / Visual and representation RL
Causal Discovery with Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RL search with BIC rather than GES with BIC | causal discovery on LiNGAM and linear-Gaussian data · 12-node LiNGAM and linear-Gaussian | SHD: 32.8\,$\pm$\,2.93 -> 17.4\,$\pm$\,7.50 (LiNGAM); 33.2\,$\pm$\,2.48 -> 22.2\,$\pm$\,6.34 (linear-Gaussian) | Helped |
| use the equal-noise-variance BIC variant RL-BIC2 instead of RL-BIC | RL causal discovery · 12-node LiNGAM and linear-Gaussian | SHD: 17.4\,$\pm$\,7.50 -> 0\,$\pm$\,0 (LiNGAM); 22.2\,$\pm$\,6.34 -> 0\,$\pm$\,0 (linear-Gaussian) | Helped |
| apply RL-BIC2 instead of NOTEARS | causal discovery · quadratic nonlinear model | SHD: 14.8\,$\pm$\,3.37 -> ~~0.6\,$\pm$\,1.20 | Helped |
| apply RL-BIC instead of GraN-DAG | causal discovery · Gaussian-process nonlinear model | SHD: 10.2\,$\pm$\,2.39 -> ~~6.2\,$\pm$\,1.33 | Helped |
| replace the single layer decoder with a bilinear decoder (ablation) | RL graph generator · Linear-Gaussian, 12-node | SHD: 0\,$\pm$\,0 -> ~~4.0\,$\pm$\,7.04 | Hurt |
| replace the single layer decoder with an NTN decoder (ablation) | RL graph generator · Linear-Gaussian, 12-node | SHD: 0\,$\pm$\,0 -> ~~7.4\,$\pm$\,8.52 | Hurt |
| replace the single layer decoder with a Transformer decoder (ablation) | RL graph generator · Linear-Gaussian, 12-node | SHD: 0\,$\pm$\,0 -> ~38.2\,$\pm$\,3.12 | Hurt |
| add quadratic regression-based pruning (ablation) | NOTEARS estimates · Quadratic nonlinear model | FDR: 0.35\,$\pm$\,0.06 -> 0.15\,$\pm$\,0.10 | Helped |
| add quadratic regression-based pruning (ablation) | NOTEARS estimates · Quadratic nonlinear model | TPR: 0.71\,$\pm$\,0.16 -> 0.70\,$\pm$\,0.15 | No clear effect |
| use NOTEARS-3 with quadratic causal functions rather than ordinary NOTEARS (ablation) | NOTEARS · Quadratic nonlinear model | SHD: 14.8\,$\pm$\,3.37 -> ~~5.2\,$\pm$\,5.19 | Helped |
| use RL-BIC2 rather than RL-BIC with differing noise variances | RL causal discovery · 12-node LiNGAM and linear-Gaussian | SHD: 14.4\,$\pm$\,7.17 -> ~~4.0\,$\pm$\,2.61 (LiNGAM); 18.8\,$\pm$\,3.43 -> ~~4.6\,$\pm$\,3.07 (linear-Gaussian) | Helped |
Closest papers by meaning
- Hierarchical Reinforcement Learning with Targeted Causal Interventions (ICML 2025 · similarity 0.90)
- Adaptive Online Experimental Design for Causal Discovery (ICML 2024 · similarity 0.90)
- Partial Structure Discovery is Sufficient for No-regret Learning in Causal Bandits (NeurIPS 2024 · similarity 0.89)
- Interventional Causal Discovery in a Mixture of DAGs (NeurIPS 2024 · similarity 0.88)
- On Causal Discovery in the Presence of Deterministic Relations (NeurIPS 2024 · similarity 0.88)
- Causal Discovery via Bayesian Optimization (ICLR 2025 · similarity 0.88)
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.