Take and win: AI wins poker tournament with four pros

Artificial intelligence (its weak form), developed by representatives of the School of Informatics at Carnegie Mellon University, beat four professional poker players. This is a Texas Hold'em tournament held at the Rivers' Pittsburgh casino. This is one of the most popular types of poker. During the tournament, about 120 thousand hands were played. AI took the bank at $ 1.7 million.

The authors of the program are Thomas Sandholm and Noam Brown. According to the developers, poker is a game whose participants do not know what cards everyone has. For this reason, it is extremely difficult to calculate. Jimmy Chu, one of the participants in the tournament, complained that he and his colleagues underestimated the capabilities of Libratus: “The bot played better and better every day. He seems to be our more skillful version. ”

The tournament in question is called " Brains vs. Artificial Intelligence: Upping the Ante ". It began on January 11th. There were five participants. One of them is the bot Libratus. The four others are Jimmy Choo, Dong Kim, Jason Les and Daniel McAulay. The total duration of the tournament was 20 days. All four people involved are considered the best players in this type of poker in the world.
')
According to the developers, winning their poker program is one of the biggest achievements in the field of artificial intelligence.


Unfortunately, the authors of the system Libratus do not reveal the features of its work. The principle of the algorithm is known only in general terms. However, in the near future technical details will be disclosed in a publication in a peer-reviewed scientific journal.

The developers claim that Libratus consists of three parts. This is the main core, a system that tracks the mistakes made by rivals and the part that tracks the weaknesses of itself. During the game, the third part helped to improve the abilities of the program, as well as improve the work of the first two parts. Libratus learned to bluff and recognize bluffs from rivals. The total computation time was approximately 15 million core hours.

Most interestingly, the Libratus algorithms are not "sharpened" specifically for poker or any other games. The program had to develop its own style of play using internal resources. And this work was based on the information provided by the developers initially - the rules of poker.

As for the prize money, the tournament participants will get $ 200,000 - this amount will be divided among all human players in accordance with the results obtained by them. Developers will not bring money home. All this was stipulated in advance, and the main thing for the scientists who created the system is the proof of its effectiveness. According to Carnegie Mellon employees, Libratus capabilities can be used in cybersecurity, in negotiations, for military purposes.



In general, artificial intelligence every year beats the person all in new games that were previously considered inaccessible to the computer. Initially these were checkers, then chess, video games, and now poker. And it is interesting that in all previous cases, all participants have complete information about the course of the game - you can see the playing field and all the chips / figures. In the latter case, the information is not complete, poker in this regard stands out.

The AI ​​has already played poker before, but it was possible to win only in the simplest version of the game - limit hold'em with a limited step of raising the stakes. But in no-limit poker everything is much more complicated. Here you need to act as based on information about your own cards, and on the data on the actions of your opponent. Moreover, opponents can bluff, which further complicates the task.

In no-limit hold'em, another software platform, DeepStack, shows excellent results. It was developed at the University of Alberta, and employees of the university, a special research group, have been creating poker bots since the 90s of the last century. The first bot created here was Loki in 1997. Then there were Poki (1999), PsOpti / Sparbot (2002), Vexbot (2003), Hyperborean (2006), Polaris (2007), Hyperborean No-Limit (2007), Hyperborean Ring (2009), Cepheus (2015) and, finally , the crown of creation - DeepStack.

Quite a lot is known about the technical capabilities of this program, since the developers published information about it in early January of this year. DeepStack uses Nash equilibrium. So called a set of different strategies in which no player can increase the gain in the event of a strategy change, if other participants in the game do not change strategies. Plus, DeepStack counteracts attempts by other players to use their own strategy.