Home > News content

Pittsburgh Poker Texas Man-made stakes showdown, AI half-way lead

via:博客园     time:2017/1/25 23:31:03     readed:1357

A man-machine in the area of ​​Texas Hold'em is being played at the Rivers Casino in Pittsburgh, United States.

The game is called "Brains Vs. Artificial Intelligence: Upping the Ante" (human brain war computer: upgraded version), in the local time on January 11 opening, will last until January 30. Four top human poker players, Jason Lee, Dong Kim, Daniel McAulay and Jimmy Chou, plus an artificial intelligence called Libratus, will compete for a total of 120,000 hand-to-head championships and 200,000 Dollar bonus.

As of January 23, which is the ninth game day, Libratus has been leading 49240 hand 459154 dollars in chips. But no matter how the results, "the first victory over human AI players rushed" this name has been less than it.

Man-machine "Huashan Mountain Sword", AI half-way lead

Libratus was developed by Tuomas Sandholm, a professor of computer science at Carnegie Mellon University in the United States, and his graduate student, Noam Brown. Sandholm is an expert in the field of game theory and artificial intelligence. He believes that under the artificial intelligence attack, the human can stick to the poker position for a long time, is very great.

This is not Sander for the first time with his AI to participate in Texas Hold'em poker man-machine war. April 24, 2016 to May 7, also in Pittsburgh's River Casino, Sandholm team developed artificial intelligence Claudico in two consecutive weeks of no-limit Hold'em in the tournament in turn with four ranked in the world's top ten Of the top poker players (Doug Polk, Dong Kim, Bjorn Li and Jason Les) singled out 80000 rounds and Claudico was fourth. The final ranking order is: the top of the Bjorn Li won a total of 529,033 US dollars, the second Doug Polk won 213,671 US dollars, the third place Dong Kim won 70,491 US dollars, while Jason Les to 80,482 US dollars behind the gap behind the AI.

And earlier in July 2015, Sandholm's Artificial Intelligence Tartanian in the computer No Limit Hold'em tournament in the pressure of the whole audience.

The play of Libratus sure than last year's defeat Claudico "deeper skill". The improvements Sander had made were mainly based on a new equilibrium approximation and several new methods of analyzing possible outcomes in later hands.

The current schedule has been half, Libratus leading four human professional players $ 460,000 chips, although not that victory in the grip, but at least is a great chance of winning.

This could be the first AI to defeat a human poker player, who had been cut off at the beginning of the month

Sandholm for the field of German man flapping "Huashan sword" rally for a long time, Ben Li Libratus to become the first victory over human players in the history of the German flutter AI, however, this honor in the The beginning of the month was unfortunate & ldquo; cut off the Hu & rdquo ;.

At the beginning of January, 10 scientists from Canada and the Czech Republic uploaded a paper entitled "DeepStack: Professional-level Artificial Intelligence Player" on the pre-printed website arXiv, introducing a technique that can be used in one-to- A New Algorithm for Defeating Human Players in Texas Hold'em. DeepStack has the "intuition" when playing cards. DeepStack learned to reason in every specific situation after applying deep learning and repeated self-gaming. This is very close to the human player "brand sense", that is, in the current context of the size of the individual card feel and make the appropriate decision. In view of Sandholm has his game schedule to tell the world, the industry generally believe that Canada and the Czech Republic in time for the team to upload papers on the eve of the start of the game, is deliberately "cut-off".

The team invited 33 professional poker players from 17 countries to challenge DeepStack, which ran 44,852 bouts between November 7, 2016 and December 12, 2016. DeepStack became the first artificial intelligence to beat a human player in one-to-one infusion, with an average winning percentage of 492mbb / g (50mb / g for professional gamers).

Although Libratus has missed the title of "history's first", Sandholm says Libratus is far more statistically significant than his opponent, who is far stronger than Deepstack and has more rounds of engagement.

Why do people persist in the poker position than Go for a long time?

Human play games, but artificial intelligence has long been nothing new. Since the Alpha dog in March 2016 to challenge world champion Li Shishi, the human field of chess has been fully occupied. So, why does poker fall after chess and where does the AI ​​mean to capture poker?

This is mainly because, after 20 years of artificial intelligence to overcome the backgammon, checkers, chess and chess are "perfect information" game, all players in the game can get certainty information is symmetrical. Artificial intelligence to master the difficulty of these games, mainly depends on the game need to make the decision points (decision points) the number, which determines the amount of computing needs of the computer. Go is the number of decision-making point in the above-mentioned several kinds of chess, and thus is finally the master of artificial intelligence. A game of chess contains about 10 to 170 power points.

And the father of computer von Neumann said, "The real world is different from this, the real world contains a lot of bets, some deceptive tactics, but also involves you will think about what others will think you will do." "Texas Hold'em is such a closer to the real world of the game, players can only master the cards in their hands, through this asymmetric information and the opponent to play.

Therefore, although the one-to-one infinity game consists of ten 160-th decision points, which is less than Go, it poses a higher requirement for the reasoning ability of artificial intelligence

MIT Technology Review commented that if Libratus were to win the event, it would be a major event for artificial intelligence. Artificial Intelligence had previously proved extremely difficult to imitate human reasoning and intelligence in the movement of poker games. Poker and checkers, chess and chess are essentially different, because poker is a "perfect information" game, in the game throughout the opponent in the hands of the cards are unknown. Therefore, it is an extremely complicated operation to give the best exit strategy based on all the possible paths of the opponent, and in the infinite set, the other party can add any note in principle, which will undoubtedly greatly increase the challenge .

AI developers use game theory and strategic decision theory to find the best strategy in a scenario full of uncertainty, the so-called "equilibrium". Because the situation involves too many possibilities, developers are more or less to use some similar techniques.

Andrew Ng, chief scientist at Baidu, said, "Poker is one of the hardest games to capture in artificial intelligence. Each step is not the so-called optimal solution, artificial intelligence to take a random strategy, so it bluff when the other party will not be allowed to eat. & Rdquo;

Bluff is a classic strategy of Texas Hold'em, which vividly embodies the "psychological game" of poker games. It is characterized by the fact that when the cards in hand are not big enough, they still raise their bravado to scare them off. opponent. In order to achieve a good bluff effect, the player's bet strategy needs to have sufficient randomness, in order to avoid the opponent to find out the routine. Always bluffing people and never bluffing people are not a good German flutter players.

Just as Kenny & Rogers wrote in the lyrics, "A gambler should know when to follow the cards and when to fold." In this man-machine battle, several human players have found that Libratus is very good at bluffing and not bluffing.

MIT's Science and Technology Review notes that the technology used to develop smarter robots will have many practical uses. Previously, game theory has been used in jamming attacks, network security, taxi autopilot, robot planning research, these studies are not perfect information in the design of decision-making context.

But even if Libratus won the championship, the human will still have a place at the table: Libratus and its representative artificial intelligence technology can not cope with the multiplayer infusion.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments