CHAI PhD student Rohin Shah and intern Micah Carroll wrote a post on human-AI collaboration on the
Berkeley AI Research Blog.
The article, titled “Collaborating with Humans Requires Understanding
Them,” is based on the paper “On the Utility of Learning about Humans for Human-AI Coordination.”
They explore whether self-play, in which AI train by playing against themselves, is a good algorithm
to use to create agents that can collaborate with human players. They use a simplified version of
the game Overcooked to test AI agents in an environment that requires challenging coordination. They
find that it is important to model humans; even using a naive human model leads to significant
improvements over self-play when playing with humans.