Jakob Foerster
TalkRL: The Reinforcement Learning Podcast
00:00
The Importance of Agent Design in Coordination
I heard about your Hanabi work, I guess last year. And so I got the game with my family and we did play anabi. And it's a very strange sensation playing that game because it's like the opposite of most games where you cannot see your own cards. The bad news is what I found pretty soon is that these agents that you train in simulation on the game from scratch are quite brittle in terms of independent training runs. So if you run the same training algorithm twice, you can get our teams that are completely incompatible. It turns out that the human ability to coordinate is quite amazing.
Play episode from 24:04
Transcript
