Research map / Reinforcement learning / Visual and representation RL
Atari-5: Distilling the Arcade Learning Environment down to Five Games
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 |
|---|---|---|---|
| Select five games using cross-validated linear regression to predict the ALE median | ALE evaluation · ALE 57-game median prediction | R²: 0.756 -> 0.984 | Helped |
| Use the Atari-5 five-game subset rather than the DQN-7 game set | ALE evaluation · ALE 57-game median prediction | approx. relative error: 38.9% -> 10.4% | Helped |
| Extend the subset from five games to ten (ablation) | Atari-5 · ALE 57-game median prediction | approx. relative error: 10.4% -> 7.2% | Helped |
| Extend the subset from five games to ten (ablation) | Atari-5 · ALE games: variance in log-scores | explained variance: 71.5% -> 80.0% | Helped |
| Reduce the selected subset from five games to three (ablation) | Atari-5 · ALE 57-game median prediction | approx. relative error: 10.4% -> 13.7% | Hurt |
| Reduce the selected subset from five games to one (ablation) | Atari-5 · ALE 57-game median prediction | approx. relative error: 10.4% -> 27.4% | Hurt |
| Use Atari-5 games to predict scores for all ALE games | Atari-5 · ALE games: variance in log-scores | explained variance: 71.5% | Mixed |
| Normalize case-study estimates to Rainbow before comparing algorithm performance | Atari-5 case-study predictions · Conjugated Discrete Distributions case study | average relative error: 18.9% -> 6.7% | Helped |
| Apply Atari-5 under a different frame time limit | Go-Explore · ALE with 400,000 frame time limits | relative overestimate: 12.2% | Hurt |
| Use Atari-5 to estimate median scores for the case-study algorithms | C51, IQN, C2D, and Rainbow · Conjugated Discrete Distributions case study | algorithm ordering: one inversion | No clear effect |
Closest papers by meaning
- On Bonus Based Exploration Methods In The Arcade Learning Environment (ICLR 2020 · similarity 0.89)
- Deep Reinforcement Learning at the Edge of the Statistical Precipice (NeurIPS 2021 · similarity 0.89)
- Agent57: Outperforming the Atari Human Benchmark (ICML 2020 · similarity 0.89)
- Investigating Pre-Training Objectives for Generalization in Vision-Based Reinforcement Learning (ICML 2024 · similarity 0.88)
- Learning to Play Atari in a World of Tokens (ICML 2024 · similarity 0.87)
- Craftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning (ICML 2024 · 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.