Finally AI defeats professional poker players in six-player poker
by
Poker is an 'incomplete information game' that fights when the opponent's hand is not revealed, and it is different from a 'complete information game' such as Go or Shogi where all information such as the opponent's game pieces is disclosed. It is said that AI ' Pluribus ' jointly developed by Facebook and Carnegie Mellon University has won the six-player poker game against such top poker pro players.
Superhuman AI for multiplayer poker | Science
https://science.sciencemag.org/content/early/2019/07/10/science.aay2400
Facebook, Carnegie Mellon build first AI that beats pros in 6-player poker
https://ai.facebook.com/blog/pluribus-first-ai-to-beat-pros-in-6-player-poker/
Humans Fold: AI Conquers Poker's Final Milestone-Scientific American
https://www.scientificamerican.com/article/ai-conquers-six-player-poker/
The research team of Tuomas Sandholm et al at Carnegie Mellon University developed an AI called ' Libratus ' in 2017 and played 120,000 games against poker professionals. As a result, it was proved that Libratus beat the professional opponent brilliantly and can win against the human opponent with high AI even in the incomplete information game.
However, the match played in 2017 was only a one-on-one match between poker and AI, and if there were more people to play with, it would be difficult for AI to win, says Sandholm. That he was thinking. Two-player poker, like Go, Shogi and Chess, has only one winner and one loser. On the other hand, in multiplayer poker, it is necessary to consider multiple decision-making processes and hands, which will increase the level of difficulty for AI.
Sandholm, who developed a new poker AI 'Pluribus' in collaboration with Facebook in 2019, two years after the match between Libratus and PokerPro, challenged the match with six players including PokerPro again .
by
Libratus and Pluribus played the rule of Texas Hold'em , the most mainstream in the world, instead of five-draw poker, which trades five cards in the mainstream in Japan for cash. Texas Hold'em combines two cards that only individual players can use with five cards that all players can use, for a total of seven cards to create a poker role.
Also, each time a card is dealt, the player performs an action to bargain with other players. The action is ' check (wait for another player's action without betting chips)' ' bet (betting a chip)' ' raise (betting more chips on another player's bet)' ' call ( There are such things as paying the same amount of chips to the bet and raise of the opponent, participating in the game, and ' folding ' without getting paid to the bet and raise of the opponent.
The player determines the strength of the role the opponent has or is about to make from the fact that the opponent has a part of the role of the opponent and the opponent's action, etc. It is possible to reflect on Because everything is not left to luck but there is room for skill to intervene, high-skilled players can achieve a higher winning rate than novice players by repeating the game over and over again.
by World Poker Tour
Pluribus sought to refine his strategy by finding his own hand, place cards, better choices for each action, and self-study through repeated copies of himself. Running 512GB of RAM using a 64-core CPU server, Pluribus learned to play poker for eight days, so the cost for training was only $ 150, even at the general cloud computing cost. Less than about 16,000 yen). The research team says training was very cheap. In practice, two CPUs and 128 GB of memory are used as resources to run Pluribus, but Alpha Go, which beat Go's top イ セ セ 九 9 levels, has 1920 CPUs and 280 It uses as many GPUs as possible, and Pluribus is characterized by its ability to run with very few resources.
At first, trained Pluribus played 10 '6 games of 5 people + Pluribus (AI)' in 10,000 days in 12 days. The poker pros chosen as the human side are top pros who have earned at least $ 1 million (about 108 million yen) in cash prizes in past poker tournaments, and 15 people rotate to join AI It was said that he played a match. Among these 15 are Greg Marson, who has won the main event of the World Series of Poker (WSOP), a poker world tournament held in the US, and side event winners. It is said that Anthony Gregg and others were included. Poker professionals were paid for each battle and were motivated to receive $ 50,000 (approximately 5,400,000 yen) based on their performance, but the result after 10,000 games was It was a great win for Pluribus.
There will also be a six-man match in the style of “one human + five Pluribus (AI)”, where Chris Ferguson who has won the WSOP Main Event and four World Poker Tour (WPT) events. Darren Elias, who has won the championship, was included in the human side. Again, the research team says Pluribus has succeeded in winning the human side and AI has been proven to be able to win six-player poker as well.
by on
Ferguson was 'pluribus very strong opponent' for Pluribus's play, 'It was difficult to identify what kind of hand he had, and he was good at bets to earn thin profits. It was also good to get a tip from the other party when I had it, 'commented. Elias also pointed out that Pluribus' strength is 'the ability to mix strategies.' If a certain tendency is seen in the play, a professional player can take action and take action, but Pluribus seems to have been able to randomly distribute the playing at a level impossible for humans.
Professional players say that Pluribus's play was astonishing to human players. Pluribus, for example, played 'Don Bet (the next action after calling and finishing the bet of the opponent's bet, bet ahead of the opponent)' much more often than a pro, but generally Donk bets are bad Is said to be. Michael Gaggliano, who played against Pluribus, said, “It was very tempting to see the poker AI and see its strategy.” “There is no human player at all, especially with regard to the bet and raise chip amounts. There were a few. '
'This is the first benchmark in a game that is neither a two player game nor a zero-sum game,' said Nome Brown of Facebook, a Pluribus co-developer. Pluribus' know-how in many places beyond poker because there are various fields such as cyber security, financial transaction, business front, etc. other than poker where the situation where the opponent's thinking is unknown can be judged The research team believes that there is a possibility that
by geralt
Related Posts:
