Jakob Foerster
TalkRL: The Reinforcement Learning Podcast
00:00
Meta Self Play: A New Thing in Meta Learning
In M for us is that we are truly model free. So and we don't we really have a meta state, right? And it allows us to have shapers that are consistent. What I mean by that is I don't have to model the other side as naive learner like in Lola. I can instead have a shaping policy that is cognizant of the fact that the other agent is also shaping at the same time. This is I believe new. I haven't come across it before.
Transcript
Play full episode
