Research map / Reinforcement learning / Offline RL
Benchmarks for Deep Off-Policy Evaluation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use model-based and direct value-based OPE rather than importance-sampling methods | DOPE RL Unplugged · DOPE RL Unplugged tasks | overall performance across metrics | Helped |
| Use FQE-L2 instead of other evaluated methods | DOPE D4RL · DOPE D4RL | absolute error and Regret@1 | Helped |
| Use IS or DR rather than FQE-L2 | DOPE D4RL · DOPE D4RL | rank correlation | Helped |
| Evaluate under the more ideal coverage and support conditions of RL Unplugged rather than D4RL | DOPE benchmark · RL Unplugged vs D4RL | rank correlation: 0.69 vs 0.25 | Helped |
| Use importance sampling instead of direct value-based methods | DOPE RL Unplugged · Humanoid run | rank correlation | Helped |
| Change the model-based architecture from feed-forward to autoregressive | Model-based OPE · DOPE RL Unplugged tasks | performance | Mixed |
| Change FQE implementation from distributional to L2 | FQE · DOPE RL Unplugged tasks | performance | Mixed |
| Use different offline datasets with varying distribution and coverage properties | OPE evaluation · D4RL Gym-MuJoCo, Adroit, and AntMaze | Regret@1 | Mixed |
| Use importance sampling rather than the other evaluated methods | OPE methods · AntMaze | rank correlation | Helped |
| Evaluate on high-dimensional tasks with temporally extended control | OPE methods · AntMaze | absolute error | Hurt |
| Use rank correlation or regret in addition to absolute error | OPE evaluation protocol · DOPE benchmark | method assessment | Mixed |
Closest papers by meaning
- OPERA: Automatic Offline Policy Evaluation with Re-weighted Aggregates of Multiple Estimators (NeurIPS 2024 · similarity 0.94)
- Active Offline Policy Selection (NeurIPS 2021 · similarity 0.92)
- Cross-Domain Off-Policy Evaluation and Learning for Contextual Bandits (ICLR 2025 · similarity 0.89)
- Model Selection for Off-policy Evaluation: New Algorithms and Experimental Protocol (NeurIPS 2025 · similarity 0.89)
- Off-Policy Evaluation under Nonignorable Missing Data (ICML 2025 · similarity 0.88)
- STITCH-OPE: Trajectory Stitching with Guided Diffusion for Off-Policy Evaluation (NeurIPS 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.