Summary of Candor: Counterfactual Annotated Doubly Robust Off-policy Evaluation, by Aishwarya Mandyam et al.
CANDOR: Counterfactual ANnotated DOubly Robust Off-Policy Evaluation
by Aishwarya Mandyam, Shengpu Tang, Jiayu Yao, Jenna Wiens, Barbara E. Engelhardt
First submitted to arxiv on: 11 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Machine Learning (stat.ML)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper proposes a family of off-policy evaluation (OPE) estimators inspired by the doubly robust (DR) principle. The authors introduce three strategies for incorporating expert-annotated counterfactual samples into a DR-inspired estimator, which can improve behavior dataset coverage. By combining importance sampling (IS) with a reward model estimate, known as the direct method (DM), these estimators offer favorable statistical guarantees. The proposed methods are evaluated in three contextual bandit environments, and empirical results show that using imperfect annotations only in the DM portion of the estimator is most beneficial when the reward model is misspecified. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Off-policy evaluation helps ensure a policy will work well before it’s used. Some experts have created ways to improve this process by adding special “counterfactual” samples. But these methods can be tricky and don’t always work well. This paper introduces new ways to use these counterfactual samples that are more reliable and accurate. The authors tested their ideas in different scenarios and found that using imperfect information only in certain parts of the process works best. |
