How Do You Determine the Data Augmentation Satisfaction Point?

We already have so much variance, like these things are super high variance anyway. You don't want more variance for no good reason. So what we basically did was that when we're training the network, we split the data set and just augment half of the data set. And then we wouldn't augment the Q targets either just to bring the variance down to reasonable levels again. That meant that we can train our policy network, our value function, we want to augment the data pretty strong argumentation. We could generalize very well to these visual randomizations that we have been smarting on for a while.