<ul><li>Distributionally robust policy learning focuses on finding a policy that performs well even under the worst-case distributional shift.</li><li>Existing methods consider the worst-case joint distribution, which can be overly conservative.</li><li>This study addresses robust policy learning under concept drifts, where the conditional relationship between outcome and covariate changes.</li><li>Proposed methods include a doubly-robust estimator and a learning algorithm for maximizing policy value, showing improvement over existing benchmarks.</li></ul>

Distributionally Robust Policy Learning under Concept Drifts

Discover more