The Meta-Rl Problems in Meta-Learning Papers

We have two axes along which we distinguish these meta-RL problems. So there's zero or few shots versus many shots. That has to do with the horizon of the task in the inner loop. And then many shots is more about learning the whole, like learning a sort of long running oral algorithm that you can update your policies 10,000 times. This could be originally these were the kinds of things where you have maybe a mojo-co environment where you have a cheetah robot running around and you need to decide which way to run with the cheetah.