Jakob Foerster
TalkRL: The Reinforcement Learning Podcast
00:00
The Nash Equilibrium of Meta Self Play
The way that meta self play works is in the end, we get these agents that are free shaping aware. They are approximately optimally shaping another shaper. So this is this recursion to us in a sense, but because we have a training process where we can anneal the probability of playing with the naive learner bus it actually gives us a specific equilibrium. That ends up being a shaping aware equilibrium that stabilize cooperation between these M for us agents.
Transcript
Play full episode
