Reproduce “Causal World Models by Unsupervised Deconfounding of Physical Dynamics” 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
Minne Li, Mengyue Yang, Furui Liu, Xu Chen, Zhitang Chen, Jun Wang
Researchers from University College London and Huawei Noah's Ark Lab developed Causal World Models (CWMs) that perform unsupervised deconfounding of physical dynamics. These models achieve superior counterfactual prediction and reduce sample complexity in model-based reinforcement learning tasks by explicitly modeling latent confounding factors.
Key insights
- Explicitly modeling latent confounding factors, even when unobserved, is crucial for accurate counterfactual prediction in dynamic physical environments.
- Unsupervised learning of both object-centric state abstractions and time-invariant confounders is achievable and enhances the robustness of world models.
- Mitigating bias from historical observational data using a doubly robust objective function significantly improves the reliability of counterfactual predictions, especially in continuous intervention settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.