John Schulman
TalkRL: The Reinforcement Learning Podcast
00:00
Using an on Policy Algorithm With a Reward Model
This is like offline RL, but you're doing it in a different way. You're using an on policy algorithm with a reward model as opposed to maybe a more typical way to do offline RL would be use off policy algorithm. Would that work here or would that not work here? What we're doing here is kind of like model based RL because the reward model is like a model of the unknown part of the system. Here is the human radar or yeah, the human, it's not the, uh, outputting, appending to your list of tokens.
Transcript
Play full episode
