Viliam Lisý presents Learning to play large imperfect-information games

On 2019-02-28 - 2019-02-28 16:15:00 at Posluchárna S5, MFF UK, Malostranské nám. 25, Praha 1
The fortieth meeting of the Prague computer science seminar

Outperforming humans in a well-defined intellectually challenging task, which
the humans spent decades practicing and studying, is a clear sign of
intelligence. Therefore, outperforming professional players of checkers, chess,
backgammon, go and poker have been important milestones in artificial
intelligence research. Solving large games is also very useful in practical
applications, for example in physical and network security.

In this talk, I will briefly introduce the key AI methods behind computing
strategies in chess and Go. Then we will focus on imperfect information games,
where players do not have the same information about the state of the game.
Approximating optimal strategies for these games is fundamentally more
and simple adaptations of the techniques from perfect information games do not
lead to good performance. I will explain the algorithm we developed for
DeepStack, the first computer program that outperformed human professionals in
no-limit Texas hold’em poker. Evaluating the performance of bots in such a
complex game presents interesting challenges and I will explain how we overcame
them. Finally, I will talk about the limitations of the algorithm used by
DeepStack and future research directions inspired by these limitations.
Responsible person: Petr Pošík