Research map / Reinforcement learning / Offline RL
Distributionally Robust Policy Evaluation and Learning in Offline Contextual Bandits
Nian Si, Fan Zhang, Zhengyuan Zhou, Jose Blanchet
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 changed | Where | Measured | Effect |
|---|---|---|---|
| Normalizing the DR objective by the propensity sum (stable version, Algorithm 3) instead of unnormalized (Algorithm 2) (ablation) | DR policy learning with depth-3 trees · Simulation worst-case KL ball test set | Q_full_DRO value: 0.075 vs 0.067 | Improved |
| Normalizing the DR objective (stable version) in policy learning (ablation) | DR policy learning · Simulation same-environment test set | Q_full_emp value: 0.243 vs 0.240 | No clear effect |
| Distributionally robust objective (stable DRO policy learning) vs non-robust empirical policy learning | DRO policy learning · Simulation worst-case KL ball test set | Q_full_DRO value: 0.075 vs 0.059 | Improved |
| Distributionally robust objective (stable DRO) vs non-robust policy learning | DRO policy learning · Simulation min over 100 KL-perturbed test sets | Q_full_min value: 0.203 vs 0.191 | Improved |
| Distributionally robust objective (stable DRO) vs non-robust policy learning in unperturbed environment | DRO policy learning · Simulation same-environment test set | Q_full_emp value: 0.243 vs 0.243 | No clear effect |
| Stable (normalized) estimator vs unnormalized estimator for fixed-policy evaluation (ablation) | DR policy evaluation · Simulation n=10^5, delta=0.2 | policy value estimate: 0.0245 vs 0.0247 | No clear effect |
| Increasing sample size from 10^3 to 10^5 for DR policy evaluation | DR policy evaluation (Algorithm 1) · Simulation delta=0.2 | standard error of estimate: ±.0312 at n=10^3 to ±.0030 at n=10^5 | Improved |
Closest papers by meaning
- Factored DRO: Factored Distributionally Robust Policies for Contextual Bandits (NeurIPS 2022 · similarity 0.92)
- Offline Neural Contextual Bandits: Pessimism, Optimization and Generalization (ICLR 2022 · similarity 0.91)
- Off-Policy Actor-Critic for Adversarial Observation Robustness: Virtual Alternative Training via Symmetric Policy Evaluation (ICML 2025 · similarity 0.89)
- Direction-Aware Offline-to-Online Learning in Linear Contextual Bandits (NeurIPS 2026 · similarity 0.89)
- Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection and Learning (NeurIPS 2024 · similarity 0.89)
- Distributionally Robust Policy Learning under Concept Drifts (ICML 2025 · similarity 0.88)
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.
