Towards Robust Off-Policy Evaluation via Human Inputs

Human expertise improves policy evaluation in high-stakes domains: The paper demonstrates that incorporating domain knowledge about which environmental factors are likely to change produces more realistic and useful policy evaluations than existing methods that assume worst-case scenarios across all possible changes. This is particularly important in healthcare, where overly pessimistic estimates could reject policies that would actually help patients.

Balancing robustness with practicality: Traditional robust evaluation methods often reject potentially valuable policies because they assume any aspect of the environment could change unpredictably. By restricting robustness analysis to plausible shifts identified by human experts, the proposed ROPE framework achieves a middle ground—policies are still evaluated conservatively, but not so pessimistically that useful interventions are discarded.

Computational efficiency matters for real-world deployment: The paper develops algorithms that are computationally efficient for both contextual bandits and Markov decision processes, addressing a practical barrier to implementing robust evaluation methods in actual healthcare systems where computational resources may be limited.

Responsible AI requires human-in-the-loop validation: The framework explicitly incorporates human judgment about which environmental changes are realistic, acknowledging that AI systems cannot determine plausibility alone and that domain experts must guide assumptions about how deployment environments will evolve.