Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Human-level Atari 200x faster
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 online-network bootstrapping with an approximate trust region instead of target-network bootstrapping without the trust region (ablation) | MEME · Atari ablation games | normalized area under score curve | Helped |
| Add the approximate trust region while retaining target-network bootstrapping (ablation) | MEME · Atari ablation games | normalized area under score curve | Helped |
| Replace Retrace with Soft Watkins Q(lambda) (ablation) | MEME · Atari ablation games, especially sparse-reward games | data efficiency | Helped |
| Add policy distillation from the value-induced policy (ablation) | MEME · Atari ablation games | convergence speed | Helped |
| Add cross-mixture training (ablation) | MEME · Atari ablation games | data efficiency | Helped |
| Use a combined loss for intrinsic and extrinsic values rather than separate losses (ablation) | MEME · Atari ablation games | data efficiency | Helped |
| Normalize TD errors for both loss and priorities (ablation) | MEME · Atari ablation games | performance | Helped |
| Replace the shared-torso separate-head representation with a single head estimating the combined intrinsic-plus-extrinsic value (ablation) | MEME · Atari ablation games | performance | No clear effect |
| Use an explicit L2 penalty between online and target networks instead of the trust region (ablation) | MEME · Atari ablation games | sample efficiency | Hurt |
| Fix the stochastic-depth mask for periods at inference as approximate Thompson sampling (ablation) | MEME · Some Atari games | performance | Mixed |
| Mix online data with replay data (ablation) | MEME · Atari, including Pitfall | stability and performance | Mixed |
| Increase samples per insert from 6 to 10 (ablation) | MEME · Atari ablation games | data efficiency | Mixed |
Closest papers by meaning
- Agent57: Outperforming the Atari Human Benchmark (ICML 2020 · similarity 0.94)
- Bigger, Better, Faster: Human-level Atari with human-level efficiency (ICML 2023 · similarity 0.90)
- Investigating Pre-Training Objectives for Generalization in Vision-Based Reinforcement Learning (ICML 2024 · similarity 0.86)
- Learning to Play Atari in a World of Tokens (ICML 2024 · similarity 0.85)
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.85)
- Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining (ICLR 2025 · similarity 0.84)
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.