Jakob Foerster
TalkRL: The Reinforcement Learning Podcast
00:00
The Meta Learning Problem in a Prison Dilemma
Chris Liu: We're defining a meta game whereby this state is augmented with the policy of the other player and each time step consists of entire episode. You have a learning you have a PPO agent that's maximizing its own return doing essentially independent learning in the prison dilemma. And then we're meta learning another PPOAgent that can learn to optimally influence knowing dynamics of this naive learner to maximize the returns of the shaper. Then as a next step, we can now train two meta agents that learn to optimality influence each other's learning process in the in this iterative game. So this is meta reinforcement learning in a specific setting and more interestingly we get to meta self play
Transcript
Play full episode
