Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels

reinforcement-learningdata-augmentationpixel-observations

Abstraction: Data augmentation technique enabling model-free RL directly from pixels

Key points:

Authors: Kostrikov, Yarats, Fergus (2020); adds input perturbations to regularize the value function in model-free RL
Applied to Soft Actor-Critic (SAC), dramatically improves training from image pixels without auxiliary losses or pre-training
Achieves state-of-the-art on DeepMind control suite, surpassing model-based methods (Dreamer, PlaNet, SLAC) and contrastive learning (CURL)
Drop-in compatible with any model-free RL algorithm requiring only minor modifications
Technique leverages augmentation strategies common in computer vision (e.g., cropping, color jitter)

Connections: Reinforcement Learning · Data Augmentation · Deep Learning