Research map / Reinforcement learning / Policy gradient and actor-critic
Off-Policy Actor-Critic with Shared Experience Replay
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 |
|---|---|---|---|
| Mix fresh on-policy experience into replay batches (ablation) | LASER with V-trace · DMLab-30 | performance | Helped |
| Use only replay data and remove online samples from the batch (ablation) | LASER with V-trace · DMLab-30 | performance | Hurt |
| Add a trust-region filter for off-policy experience (ablation) | shared-replay LASER · DMLab-30 hyper-parameter sweeps | performance | Helped |
| Share replay across agents without a trust region (ablation) | LASER · DMLab-30 hyper-parameter sweeps | performance | Hurt |
| Combine trust-region estimation with population-based training (ablation) | shared-replay LASER · DMLab-30 | performance | Helped |
| Increase replay capacity (ablation) | LASER with 87.5% replay data per batch · DMLab-30 | data efficiency | Helped |
| Increase the V-trace clipping threshold to reduce clipping (ablation) | shared-replay LASER · DMLab-30 hyper-parameter sweeps | performance | Hurt |
| Use uniform replay instead of prioritized replay (ablation) | actor-critic replay agent · DMLab-30 | performance | No clear effect |
| Use prioritized replay without recomputing LSTM states (ablation) | V-trace replay agent · DMLab-30 | data efficiency | No clear effect |
| Share experience replay across agents (ablation) | LASER · Atari 57 games, 200M environment steps | median human-normalized score: 431% -> 448% | Helped |
| Share experience replay across agents (ablation) | LASER · DMLab-30, 10B environment steps | median: 95.4% -> 97.2% | Helped |
Closest papers by meaning
- Sample Efficient Actor-Critic with Experience Replay (ICLR 2017 · similarity 0.92)
- Actor Prioritized Experience Replay (Abstract Reprint) (AAAI 2024 · similarity 0.91)
- Shared Experience Actor-Critic for Multi-Agent Reinforcement Learning (NeurIPS 2020 · similarity 0.89)
- Turning Sand to Gold: Recycling Data to Bridge On-Policy and Off-Policy Learning via Causal Bound (NeurIPS 2025 · similarity 0.88)
- Offline Actor-Critic Reinforcement Learning Scales to Large Models (ICML 2024 · similarity 0.87)
- Offline-Boosted Actor-Critic: Adaptively Blending Optimal Historical Behaviors in Deep Off-Policy RL (ICML 2024 · 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.