Style-Agnostic Reinforcement Learning

AbstractWe present a novel method of learning style-agnostic representation using both style transfer and adversarial in the reinforcement framework. The style, here, refers to task-irrelevant details such as color background images, where generalizing learned policy across environments with different styles is still challenge. Focusing on representations, our trains actor diverse image generated from an inherent perturbation generator, which plays min-max game between without demanding expert knowledge for data augmentation or additional class labels training. We verify that achieves competitive better performances than state-of-the-art approaches Procgen Distracting Control Suite benchmarks, further investigate features extracted model, showing model captures invariants less distracted by shifted style. code available at https://github.com/POSTECH-CVLab/style-agnostic-RL.KeywordsReinforcement learningDomain generalizationNeural transferAdversarial