Research map / Reinforcement learning / Policy gradient and actor-critic
Live in the Moment: Learning Dynamics Model Adapted to Evolving Policy
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 policy-distance-adjusted historical-policy weights in model learning and rollout initial-state sampling | MBPO · Hopper | asymptotic return: 3125.56 -> 3641.07 | Helped |
| Use policy-distance-adjusted historical-policy weights in model learning and rollout initial-state sampling | MBPO · Walker2d | asymptotic return: 4366.37 -> 5304.42 | Helped |
| Use policy-distance-adjusted historical-policy weights in model learning and rollout initial-state sampling | MBPO · Humanoid | asymptotic return: 4148.15 -> 5885.14 | Helped |
| Use policy-distance-adjusted weights for model learning only (ablation) | MBPO · Hopper and Walker2d | return | Helped |
| Use policy-distance-adjusted weights for rollout initial-state sampling only (ablation) | MBPO · Hopper and Walker2d | return | Helped |
| Use adjusted policy weights for both model learning and rollout initial-state sampling (ablation) | MBPO · Hopper and Walker2d | return | Helped |
| Increase current-policy proportion rate beyond the recommended small rate (ablation) | PDML-MBPO · Hopper | return | Hurt |
| Use exponential-decay policy weights instead of PDML weights | MBPO · Hopper | asymptotic return: 3125.56 -> 3320.04; 3125.56 -> 3291.93; 3125.56 -> 3382.63; 3125.56 -> 3374.52 | Mixed |
| Use exponential-decay policy weights instead of PDML weights | MBPO · Walker2d | asymptotic return: 4366.37 -> 4609.29; 4366.37 -> 4571.64; 4366.37 -> 4643.15; 4366.37 -> 4595.31 | Mixed |
| Use exponential-decay policy weights instead of PDML weights | MBPO · Humanoid | asymptotic return: 4148.15 -> 5070.58; 4148.15 -> 5198.34; 4148.15 -> 5092.56; 4148.15 -> 5149.71 | Mixed |
| Train on the adjusted policy mixture until convergence rather than the standard MBPO mixture | Dynamics model · Walker2d | one-step prediction error: 0.946 ±0.201 -> 0.673 ± 0.117; 0.578 ± 0.029 -> 0.422 ± 0.169; 0.328 ± 0.037 -> 0.207 ± 0.072; 0.317 ± 0.013 -> 0.128 ± | Helped |
| Train on the adjusted policy mixture rather than the standard MBPO mixture | Dynamics model · HalfCheetah | one-step prediction error: 0.478 ± 0.027 -> 0.363 ± 0.066; 0.345 ± 0.088 -> 0.232 ± 0.053; 0.268 ± 0.053 -> 0.206 ± 0.163; 0.245 ± 0.051 -> 0.149 ± | Helped |
Closest papers by meaning
- Prioritized Model Experience Replay (ICML 2026 · similarity 0.91)
- On-Policy Model Errors in Reinforcement Learning (ICLR 2022 · similarity 0.90)
- Policy-conditioned Environment Models are More Generalizable (ICML 2024 · similarity 0.89)
- Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning (ICLR 2025 · similarity 0.88)
- Reward-Consistent Dynamics Models are Strongly Generalizable for Offline Reinforcement Learning (ICLR 2024 · similarity 0.88)
- Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control (ICLR 2025 · similarity 0.87)
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.