I think that learning to acquire new/additional training data would be a better first step towards learning agents, than trying to mutate its structure/hyper-pa

I think that learning to acquire new/additional training data would be a better first step towards learning agents, than trying to mutate its structure/hyper-parameters.