Research map / Reinforcement learning / Visual and representation RL
Efficient Scheduling of Data Augmentation for Deep 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 |
|---|---|---|---|
| Postpone augmentation distillation until after RL training (UCB-ExDA vs. UCB-InDA) | UCB-based scheduling on Procgen · Heist (test-bg) | score: 6.23$\pm$ 1.29 -> \bf{7.86}$\pm$ 0.83 | Helped |
| Postpone augmentation distillation until after RL training (UCB-ExDA vs. UCB-InDA) | UCB-based scheduling on Procgen · Chaser (test-bg) | score: 1.08$\pm$ 0.08 -> \bf{3. 18}$\pm$ 0.79 | Helped |
| Use UCB to choose augmentation or no augmentation during RL, then distill all augmentations afterward | PPO on Procgen · Heist (test-bg) | score: 5.18 $\pm$ 1.53 -> \bf{7.86}$\pm$ 0.83 | Helped |
| Use UCB to choose augmentation or no augmentation during RL, then distill all augmentations afterward | PPO on Procgen · Chaser (test-lv) | score: 4.83$\pm$ 0.88 -> 6.43$\pm$ 1.28 | Helped |
| Apply random-color augmentation through ExDA rather than InDA | PPO on Procgen · Heist easybg (test-bg) | normalized score: 1.92 -> 2.11 | Helped |
| Apply random-crop augmentation through InDA rather than ExDA | PPO on Procgen · Heist easybg (test-lv) | normalized score: 1.09 -> 1.80 | Helped |
| Use InDA rather than ExDA with random crop | PPO on Procgen · Heist easybg (train) | normalized score: 0.91 -> 1.25 | Helped |
| Add the option of no augmentation to UCB-InDA (ablation) | UCB-InDA on Heist · Heist easybg (train) | score | Helped |
| Preserve the original policy response during augmented-observation distillation (DA rather than DrAC-style self-distillation) (ablation) | Augmentation distillation on Procgen · Procgen, six-game average | normalized score | Helped |
| Separate PPO and distillation into alternating phases rather than optimize both simultaneously (ablation) | Augmentation-aware PPO · Procgen | train performance | Helped |
| Add random black-image augmentation during RL training (ablation) | InDA on Procgen · Heist | score: 9.2 $\pm$ 0.46 -> 6.72$\pm$0.419 | Hurt |
| Add random black-image augmentation during RL training (ablation) | DrAC on Procgen · Chaser | score: 5.63$\pm$ 1.12 -> 1.49$\pm$0.036 | Hurt |
Closest papers by meaning
- Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates (ICLR 2024 · similarity 0.89)
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 2024 · similarity 0.88)
- MAD-TD: Model-Augmented Data stabilizes High Update Ratio RL (ICLR 2025 · similarity 0.87)
- How Ensembles of Distilled Policies Improve Generalisation in Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.86)
- Behavior Injection: Preparing Language Models for Reinforcement Learning (NeurIPS 2025 · 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.