AI Paper Atlas

Research map / Reinforcement learning / Policy gradient and actor-critic

Trustworthy Policy Learning under the Counterfactual No-Harm Criterion

Haoxuan Li, Chunyuan Zheng, Yixiao Cao, Zhi Geng, Yue Liu, Peng Wu

ICML 2023 · Reinforcement learning › Policy gradient and actor-critic · approach: Confounding-robust policy evaluation

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add counterfactual no-harm constraint (uFNA upper bound) to policy learningCATE-based policy learning on IHDP and JOBS · IHDP and JOBS, cost 0-0.10true harm: Naive 19.12 vs No-Harm (u) 10.34 on IHDP (c=0)Improved
add counterfactual no-harm constraint (wFNA upper bound) to policy learningCATE-based policy learning on IHDP and JOBS · IHDP and JOBS, cost 0-0.10true harm: Naive 19.12 vs No-Harm (w) 5.98 on IHDP (c=0)Improved
use the tighter uFNA bound instead of wFNA bound in the harm constraint (ablation)No-Harm policy learning · IHDP and JOBSreward: No-Harm (u) reward 496.93 vs No-Harm (w) 459.80 on IHDP (c=0); 1453.00 vs 1325.00 on JOBSImproved
switch from CATE-based to recommendation-based utility without harm constraintnaive policy learning · IHDPtrue harm: Naive CATE 19.12 vs Naive recommendation 64.36 (c=0)Hurt
increase cost c(x) from 0 to 0.10 in harm-constrained policies (ablation)No-Harm policy learning · IHDP and JOBStrue harm and reward: Reward and harm decrease with cost; welfare change does not change muchMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy