Research map / Reinforcement learning / Visual and representation RL
Episodic Memory-Guided Controllable Experience Synthesis for Reinforcement Learning
Xiao Ma, Tian Li, Wu-Jun Li
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 |
|---|---|---|---|
| condition the diffusion model on both h(s) and EM-estimated return r + gamma e(h(s')) rather than either alone (ablation) | EMCES · Hopper medexp, TD3+BC | normalized score: 91.7 / 91.0 -> 105.5 | Improved |
| use EM-estimated r + gamma e(h(s')) instead of discounted return d(s) as condition (ablation) | EMCES · Hopper medexp, TD3+BC | normalized score: 100.9 -> 105.5 | Improved |
| use EMTD-error prioritized condition sampling instead of uniform (ablation) | EMCES · Hopper medexp, TD3+BC | normalized score: 91.6 -> 105.5 | Improved |
| use EMTD-error sampling instead of return-based or reward-based sampling (ablation) | EMCES · Hopper medexp, TD3+BC | normalized score: 95.2 (return) / 99.4 (reward) / 103.6 (e(h(s))) -> 105.5 | Improved |
| use hashing-based state representation instead of grid-based (ablation) | EMCES · Hopper medexp, TD3+BC | normalized score: 98.2 -> 105.5 | Improved |
| use hashing-based state representation instead of random projection (ablation) | EMCES · Hopper medexp, TD3+BC | normalized score: 104.0 -> 105.5; storage 7806.49 -> 0.96 KB, time 7573.88 -> 297.12 s | No clear effect |
| replace original data with EMCES synthetic data of same size | TD3+BC · D4RL gym MuJoCo sum | sum of normalized scores: 683.5 (SynthER) -> 731.4 | Improved |
| use EMCES synthetic data instead of SynthER for IQL | IQL · D4RL hopper-medium-replay | normalized score: 42.3 -> 101.0 | Improved |
| EMCES synthetic data vs original dataset for EDAC | EDAC · D4RL gym MuJoCo sum | sum of normalized scores: 837.7 (original) vs 837.0 | No clear effect |
| use EMCES synthetic data on vision-based VD4RL (latent space) | DrQ+BC · VD4RL walker-walk and cheetah-run average | average episode reward: 472.38 -> 518.00 | Improved |
| use SAC (EMCES) instead of SAC (SynthER) and SAC (PGR) in online RL | SAC · online DMC and Gym (6 environments) | mean average return | Improved |
Closest papers by meaning
- Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates (ICLR 2024 · similarity 0.87)
- Prioritized Generative Replay (ICLR 2025 · similarity 0.86)
- Energy-Guided Diffusion Sampling for Offline-to-Online Reinforcement Learning (ICML 2024 · similarity 0.86)
- Open the Black Box: Step-based Policy Updates for Temporally-Correlated Episodic Reinforcement Learning (ICLR 2024 · similarity 0.86)
- ART for Diffusion Sampling: A Reinforcement Learning Approach to Timestep Scheduling (NeurIPS 2026 · similarity 0.86)
- Adding Conditional Control to Diffusion Models with Reinforcement Learning (ICLR 2025 · similarity 0.86)
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.
