Research map / Reinforcement learning / Visual and representation RL
Behavior From the Void: Unsupervised Active Pre-Training
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 particle-based entropy exploration in abstract representation space | DrQ · DMControl suite | episodic return and data efficiency | Helped |
| use particle-based entropy pre-training | DrQ · Atari 26-game subset | median HNS: 26.8 -> 47.50 | Helped |
| use particle-based entropy pre-training | DrQ · Atari 26-game subset | mean HNS: 35.7 -> 69.55 | Helped |
| use particle-based entropy pre-training | VISR · Atari 57-game suite | median HNS: 9.41 -> 33.41 | Helped |
| use particle-based entropy pre-training | VISR · Atari 57-game suite | mean HNS: 68.42 -> 47.73 | Hurt |
| add contrastive representation learning (ablation) | APT · Atari 26-game subset | median HNS: 21.23 -> 47.50 | Helped |
| add contrastive representation learning (ablation) | APT · Atari 57-game suite | median HNS: 28.65 -> 33.41 | Helped |
| use two rather than one gradient step per environment step (ablation) | APT · Atari 26-game subset | median HNS: 41.50 -> 47.50 | Helped |
| use ELU instead of the DrQ activation function (ablation) | APT · Atari 26-game subset | median HNS: 45.71 -> 47.50 | Helped |
| use k-nearest neighbors rather than randomly selected neighbors for entropy (ablation) | APT · Atari 26-game subset | median HNS: 20.80 -> 47.50 | Helped |
| learn the encoder rather than keep it randomly initialized (ablation) | APT · Atari 26-game subset | median HNS: 33.24 -> 47.50 | Helped |
| use particle-based entropy reward rather than a positive do-not-die signal (ablation) | APT · Atari Ms Pacman pre-training | unique RAM states visited: nearly 27 times more | Helped |
Closest papers by meaning
- APS: Active Pretraining with Successor Features (ICML 2021 · similarity 0.91)
- Investigating Pre-Training Objectives for Generalization in Vision-Based Reinforcement Learning (ICML 2024 · similarity 0.88)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.88)
- Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration (ICML 2025 · similarity 0.88)
- ACE: Off-Policy Actor-Critic with Causality-Aware Entropy Regularization (ICML 2024 · similarity 0.88)
- Efficient Reinforcement Learning Through Adaptively Pretrained Visual Encoder (AAAI 2025 · similarity 0.88)
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.