“Building cooperating and coordinating agents”

Speaker: Alex Peysakhovich, Facebook AI Research (FAIR)

Talk summary: Recent work in artificial intelligence has led to agents that surpass human performance in zero-sum games. However, most real life environments are not zero-sum - rather they involve cooperation and coordination. In this talk I will discuss recent work from our lab on how insights from simple matrix games (eg. the Prisoner’s Dilemma, the Stag Hunt) can be scaled to more complex environments. We will show how to construct agents that can maintain cooperation via a generalized version of tit-for-tat with both perfect and imperfect information. We will discuss how giving agents prosociality can improve payoffs in coordination games. Finally, we will discuss recent work on agents that can learn social conventions quickly.

Relevant pre-prints:

https://arxiv.org/abs/1707.01068

https://arxiv.org/abs/1709.02865

https://arxiv.org/abs/1710.06975

Hosted in partnership with the Human Cooperation Lab.