Augmented Random Search

• The authors point out that the sub-policies learned from AutoAugment are inherently flawed because of the discrete search space.
• They convert the probability and magnitude of augmentations into a continuous space and search for sub-policies with ARS.