Research map / Reinforcement learning / Visual and representation RL
Automatic Data Augmentation for Generalization in Reinforcement Learning
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 |
|---|---|---|---|
| Add actor and value-function invariance regularization to augmentation training | PPO with the best augmentation (DrAC vs RAD) · Procgen train | PPO-Normalized Return (%) mean: 109.1 -> 119.6 | Helped |
| Add actor and value-function invariance regularization to augmentation training | PPO with the best augmentation (DrAC vs RAD) · Procgen test | PPO-Normalized Return (%) mean: 131.3 -> 138.1 | Helped |
| Select the augmentation online with UCB rather than an RL2 selection policy | DrAC · Procgen test | PPO-Normalized Return (%) mean: 105.3 -> 139.7 | Helped |
| Select the augmentation online with UCB rather than meta-learning convolutional augmentation weights | DrAC · Procgen test | PPO-Normalized Return (%) mean: 101.2 -> 139.7 | Helped |
| Add policy regularization to value-function-only regularization (ablation) | DrC (forming DrAC) · Procgen test | PPO-Normalized Return (%) mean: 115.4 -> 138.1 | Helped |
| Add value-function regularization to policy-only regularization (ablation) | DrA (forming DrAC) · Procgen test | PPO-Normalized Return (%) mean: 126.6 -> 138.1 | Helped |
| Replace UCB selection with uniform random augmentation selection (ablation) | UCB-DrAC · Procgen test | PPO-Normalized Return (%) mean: 103.4 -> 139.7 | Hurt |
| Use crop augmentation for every game instead of selecting by task (ablation) | UCB-DrAC · Procgen test | PPO-Normalized Return (%) mean: 132.7 -> 139.7 | Helped |
| Remove DrAC regularization while retaining UCB augmentation selection (ablation) | UCB-DrAC · Procgen test | PPO-Normalized Return (%) mean: 125.9 -> 139.7 | Helped |
| Add DrAC regularization to UCB-selected augmentation training (ablation) | UCB-RAD (forming UCB-DrAC) · Procgen test | PPO-Normalized Return (%) mean: 125.9 -> 139.7 | Helped |
| Use UCB-DrAC rather than PPO | PPO · Procgen | JSD mean: 0.25 -> 0.16 | Helped |
| Use UCB-DrAC rather than PPO | PPO · Procgen | Cycle-consistency 2-way mean (%): 20.50 -> 27.70 | Helped |
Closest papers by meaning
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 2024 · similarity 0.91)
- Efficient Scheduling of Data Augmentation for Deep Reinforcement Learning (NeurIPS 2022 · similarity 0.90)
- Stabilizing Deep Q-Learning with ConvNets and Vision Transformers under Data Augmentation (NeurIPS 2021 · similarity 0.90)
- Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates (ICLR 2024 · similarity 0.89)
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.87)
- Reinforcement Learning with Euclidean Data Augmentation for State-Based Continuous Control (NeurIPS 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.