Research map / Reinforcement learning / Offline RL
Double Check Your State Before Trusting It: Confidence-Aware Bidirectional Offline Model-Based Imagination
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 |
|---|---|---|---|
| add CABI double-checked imagined data to BCQ | BCQ · Adroit tasks | total normalized score: 476.7 -> 607.0 | Helped |
| add CABI double-checked imagined data to TD3_BC | TD3_BC · MuJoCo tasks | total normalized score: 979.3 -> 1020.7 | Helped |
| add the double-check filter to bidirectional imagination (ablation) | BCQ on Adroit · Adroit tasks | total normalized score: 529.2 -> 607.0 | Helped |
| use forward-only rather than double-checked bidirectional imagination (ablation) | BCQ · Adroit tasks | total normalized score: 607.0 vs. 520.6 | Hurt |
| use backward-only rather than double-checked bidirectional imagination (ablation) | BCQ · Adroit tasks | total normalized score: 607.0 vs. 526.5 | Hurt |
| remove the double-check filter and admit all bidirectional synthetic samples (ablation) | BCQ · Adroit pen-cloned and pen-human | normalized average score: pen-cloned: 54.7 vs. 43.4; pen-human: 75.1 vs. 49.6 | Hurt |
| replace CVAE rollout policy with a random policy (ablation) | CABI+BCQ · four Adroit tasks | normalized average score: pen-cloned: 54.7 vs. 37.6; pen-expert: 127.6 vs. 106.3; hammer-cloned: 4.3 vs. 0.3; hammer-expert: 128.9 vs. 103.8 | Hurt |
| replace double-check selection with random selection of 20% (ablation) | CABI+BCQ · four Adroit tasks | normalized average score: pen-cloned: 54.7 vs. 41.2; pen-expert: 127.6 vs. 112.6; hammer-cloned: 4.3 vs. 0.9; hammer-expert: 128.9 vs. 104.2 | Hurt |
| select transitions by lowest ensemble variance instead of bidirectional agreement (ablation) | CABI+BCQ · four Adroit tasks | normalized average score: pen-cloned: 54.7 vs. 40.4; pen-expert: 127.6 vs. 118.8; hammer-cloned: 4.3 vs. 0.4; hammer-expert: 128.9 vs. 125.5 | Hurt |
| vary the retained fraction to k=0 or k=100 instead of k=20 (ablation) | CABI+TD3_BC · MuJoCo random and medium datasets | normalized average score: halfcheetah-random: 10.2 -> 15.1 (k=20), 11.4 (k=100); hopper-medium: 99.5 -> 100.4 (k=20), 3.1 (k=100) | Mixed |
Closest papers by meaning
- Beyond OOD State Actions: Supported Cross-Domain Offline Reinforcement Learning (AAAI 2024 · similarity 0.90)
- A Clean Slate for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Temporal Distance-aware Transition Augmentation for Offline Model-based Reinforcement Learning (ICML 2025 · similarity 0.89)
- Constrained Latent Action Policies for Model-Based Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Are Expressive Models Truly Necessary for Offline RL? (AAAI 2025 · similarity 0.89)
- Offline-Boosted Actor-Critic: Adaptively Blending Optimal Historical Behaviors in Deep Off-Policy RL (ICML 2024 · similarity 0.89)
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.