Causal Imitability Under Context-Specific Independence Relations
TL;DR.
Context-specific independences can make expert behavior causally imitable even when classical causal imitation says it is impossible. We characterize this richer decision problem, prove it is NP-hard, derive a graphical criterion, and give a sound algorithm that combines CSI relations with observed data.
Why this matters.
A small amount of structural information that only holds in particular contexts can unlock imitation that is otherwise impossible under hidden confounding. This shows that fine-grained causal knowledge can be operationally valuable for imitation learning.
Abstract
Drawbacks of ignoring the causal mechanisms when performing imitation learning have recently been acknowledged. Several approaches both to assess the feasibility of imitation and to circumvent causal confounding and causal misspecifications have been proposed in the literature. However, the potential benefits of the incorporation of additional information about the underlying causal structure are left unexplored. An example of such overlooked information is context-specific independence (CSI), i.e., independence that holds only in certain contexts. We consider the problem of causal imitation learning when CSI relations are known. We prove that the decision problem pertaining to the feasibility of imitation in this setting is NP-hard. Further, we provide a necessary graphical criterion for imitation learning under CSI and show that under a structural assumption, this criterion is also sufficient. Finally, we propose a sound algorithmic approach for causal imitation learning which takes both CSI relations and data into account.
The paper at a glance
An informal guide to the problem, the idea, and the main results.
1. The problem
Observed expert behavior can fail to identify an imitable policy when latent structure or selection effects hide the relevant mechanisms. Classical imitation-learning views can therefore be fundamentally too optimistic.
2. The key idea
We bring context-specific independence relations into the imitability question. These richer structural constraints create information that ordinary conditional independences miss.
3. Improvability through CSI
We prove that deciding imitability under context-specific independences is NP-hard, give a necessary graphical criterion, and show that the criterion becomes sufficient under a structural condition on the context variables.
4. A sound imitation algorithm
We develop a sound algorithm that combines CSI relations with observational data to construct an imitating policy when the required causal quantities can be recovered.
5. How it works
- Formalize the notion of causal imitability.
- Identify where hidden structure blocks imitation under weaker assumptions.
- Introduce context-specific independence relations to recover extra leverage.
- Use those relations to characterize when imitation becomes possible.
What the experiments show
Where this helps
Good fit
- You care about imitation learning in environments with latent causal structure.
- You have reason to believe context-specific independences are scientifically justified.
- You want to know whether failure is algorithmic or fundamentally causal.
Keep in mind
- CSI relations enrich the causal information available to imitation learning without requiring a fully observed causal system.
- The method combines structural information and data, so the extra causal knowledge directly changes which expert policies can be reproduced.
Cite this paper
@inproceedings{jamshidi2023imitability,
title = {Causal Imitability Under Context-Specific Independence Relations},
author = {Fateme Jamshidi and Sina Akbari and Negar Kiyavash},
booktitle = {Advances in Neural Information Processing Systems},
volume = {36},
year = {2023},
url = {https://proceedings.neurips.cc/paper_files/paper/2023/hash/54e13b23fa2f399cea6e67acf9063c40-Abstract-Conference.html}
}