Research map / Reinforcement learning / Visual and representation RL
High-Throughput Synchronous Deep RL
Iou-Jen Liu, Raymond Yeh, Alexander Schwing
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 |
|---|---|---|---|
| Use HTS-RL rather than synchronous A2C/PPO | deep RL training · Atari and GFootball | training speed: 2-6× faster | Improved |
| Use HTS-RL instead of A2C | Atari · Atari BankHeist | average evaluation reward: 775 ± 166 -> 942 ± 100 | Improved |
| Use HTS-RL instead of PPO | GFootball · 3 vs 1 with keeper | average score: 0.20±0.09 -> 0.81±0.02 | Improved |
| Increase actor count from 1 to 4 (ablation) | HTS-RL · GFootball 3 vs 1 with keeper | steps per second: 1229 -> 1362 | Improved |
| Increase actor count from 4 to 8 or 16 (ablation) | HTS-RL · GFootball 3 vs 1 with keeper | steps per second: 1362 -> 1393; 1362 -> 1388 | No clear effect |
| Increase actor count from 1 to 16 (ablation) | HTS-RL · GFootball 3 vs 1 with keeper | average score: 0.82 -> 0.82 | No clear effect |
| Increase synchronization interval from 4 to 512 (ablation) | HTS-RL · GFootball 3 vs 1 with keeper | steps per second: 445 -> 1377 | Improved |
| Vary synchronization interval from 4 to 512 (ablation) | HTS-RL · GFootball 3 vs 1 with keeper | average score: 0.82; 0.81; 0.81; 0.82; 0.81; 0.83 | No clear effect |
| Use one-step delayed gradient instead of truncated importance sampling (ablation) | HTS-RL · Atari BankHeist, Breakout, and Seaquest | average episode reward: 987 vs 881; 415 vs 390; 1831 vs 1827 | Improved |
| Use one-step delayed gradient instead of no correction (ablation) | HTS-RL · Atari BankHeist, Breakout, and Seaquest | average episode reward: 987 vs 877; 415 vs 402; 1831 vs 1784 | Improved |
| Train 3 agents instead of 1 | HTS-RL (PPO), raw-image input · GFootball 3 vs 1 with keeper | average game score: 0.30± 0.11 -> 0.63± 0.15 | Improved |
Closest papers by meaning
- SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores (ICLR 2024 · similarity 0.86)
- Staggered Environment Resets Improve Massively Parallel On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- MAD-TD: Model-Augmented Data stabilizes High Update Ratio RL (ICLR 2025 · similarity 0.85)
- Cleanba: A Reproducible and Efficient Distributed Reinforcement Learning Platform (ICLR 2024 · similarity 0.85)
- Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training (NeurIPS 2025 · similarity 0.85)
- AREAL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning (NeurIPS 2025 · similarity 0.85)
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.
