Thore Graepel (Google DeepMind)
David Brower Center, 2150 Allston Way, Berkeley
Intelligence can be viewed as the ability of agents to achieve goals in a wide range of environments. If we wish to use machine learning to train intelligent agents, we need ways of creating rich environments that provide appropriate challenges and feedback signals to learning agents. Just as in real life (and evolution), the most challenging environments for learning agents arise from interaction with other co-adapting learning agents. So, let's play games with AI!
The first example is learning from self-play in the context of the AlphaGo project which led to the first computer program to beat a top professional Go player at the full-size game of Go. Similar ideas can be used to study the age-old question of how cooperation arises among self-interested agents. Finally, we look at training artificial agents to play the game of Capture-The-Flag, a competitive team game played from a first-person perspective in a complex 3D world.
Registration is optional, but space is limited. Register to reserve a seat.
Theoretically Speaking is a lecture series highlighting exciting advances in theoretical computer science for a broad general audience. Events are held at the David Brower Center in Downtown Berkeley, and are free and open to the public. No special background is assumed.
Light refreshments will be served before the lecture, at 5:30 pm.