An AI beats a table professional Texas hold’em


Two IA distinct , had already managed to beat the professionals in a variant for two players there are more than two years ago, but this is the first time that an artificial intelligence won a game with more than two elite players. The fact of passing directly to six players represents a huge leap forward. In addition to adding complexity, “gross” linked to the multiplication of possibilities, this introduces a whole extra dimension, because of all the issues that intangible that is the result of interactions between players.


These tests took place over a period of 12 days compared to twelve professionals to be part of the elite world, during which more than 10,000 hands have been played according to two different configurations. In the first, Pluribus against five human players, while in the second, five versions of the bot (which could, however, not work) were playing with a human.
In both cases, the verdict is in : after one of the fathers of the project interviewed by The Verge, Pluribus has earned an average of$ 5 per hand and about$ 1000 winnings at the time, be a “decisive margin decisive victory” witness a “superhuman”. This is a view shared by Chris Ferguson, champion at the World Series of Poker and a member of the group of 12 players beaten by the AI.



Pluribus is a very tough opponent to face. It is very difficult to jam, regardless of his hand.



Texas Hold’em, a Mount Everest of the AI


This is not the first time that an AI beats a professional at his own game, far from it. Since the defeat of Garry Kasparov against Deep Blue, on February 10, 1996, the machines have started collecting the victories of this scale. Since this historic date, computers have evolved to the point of constitutes a category of its own, very largely unaffordable even for the great masters to the most talented. At this point they even have now their own competition, simply called Computer Chess Championship.


Deep Blue – © IBM

In 2018, this is another remedy that has been achieved : Lee Sedol, true god from south korea game Go flying yet the competition for almost ten years, is beaten by AlphaGo. This AI is designed by the company DeepMind Google has achieved a real feat. Until then, we thought the game go out of the reach of artificial intelligence, the fault of a number of combinations, mind-blowing, still much higher than in chess. It is estimated that there are more possibilities in the game of go than there are atoms in the universe… which, however, has not prevented the defeat of Lee Sedol.


More recently, we even witnessed the AI DeepMind ridicule the players of Dota and Starcraft II. Recently, we even learned that this same AI would attack in ladder 1V1 eu. These victories have contributed to confirm the impression which had existed for some time already : today, the bots of DeepMind are invincible by a human in games two-player games zero-sum (where one player wins exactly what the other loses, and vice versa).


But with the poker to six, the theory of games is not enough because we are dealing with a scenario where multiple parties with different interests confront each other in a game without a clear condition to defeat or victory… plus, poker is a game of incomplete information, where the player has never all of the information that would be needed to secure his victory.



This requires, therefore, to create models of decision-making of adversaries in addition to its own strategy – the two being interdependent. A level of complexity much higher, well summarized in this study. The traditional approach, which is to browse huge decision trees weighted, is so little adapted to this scenario…


Pluribus has therefore used a different approach : in contrast to the IAs to be more traditional, he does not seek the best hit by checking all the possibilities until the end of the game. He tries a couple of moves in advance, privileging the capacity ofshort-term adaptation : this approach, which might seem counter-intuitive, which proved to be decisive to be able to perform these feats. This approach also makes the AI much faster to train, and requires a hardware ridiculous in comparison to the supercomputers that usually carry out these tasks by the traditional method… which has helped the cause in only eight days for an estimated price of$150, compared to several hundreds of thousands of euros for a supercomputer !


The AI is able to search and find, rest to teach him how to invent


And the result has been quite traumatic for the players, who have all praised its “constance relentless”, especially in the bluff, making it incredibly unpredictable-continuously. And this, we thought this ability to be the own of the Man… A bit like Go, which they said he had a “human spirit” to be able to master the game. After the defeat of Sedol, it is a different conception of the human bites the dust against the AI.



Gold Pluribus has resulted in all alone against himself, from scratch, without relying on humans. As AlphaZero chess at his time or AlphaGo. This has had a result very interesting : the AI uses successful techniques are considered unreliable, or bad by the human.


Thus, in the above video, we see the commentators, and then the champion completely dazed by the 37th shot of the machine : all the experts agreed that there was “not a single human player who would have chosen this move”. It will prove yet to be a real stroke of genius, of those who are so bright that they shut down a game and remain in the annals. In the same way, Pluribus has distinguished itself by making use of strategies for unknown, or never used by humans.



And it is certainly there that the major part of the interest of this great victory. As if owning an AI capable of slaying everyone in poker is very interesting, it is clear that the practical applications are very limited. The interest lies in the fact thatextrapolate some of the mechanisms formalised by the AI during his self-learning, free of any influence (and, therefore, the prejudices and biases of reasoning) humans ! For example, if an AI is able to decide to bluff if it considers this choice to be the most judicious, we can imagine that it could, for example, be applied to areas such as the commercial negotiation, the detection of fraud or counterfeit, the real-time decisions of autonomous cars


The challenge lies in the fact that to succeed in “extracting” principles broader example, very accurate as is the poker. It’s a safe bet that the next step in this direction is to design an AI capable of reaching levels that are as superhuman in several games at once, which is part of a “general ability rather than a skill niche”. Or even, designing new games, which seems to appear as the new Everest of the AI applied to games.


© Wiki Commons

And if this obstacle is climbed, it’s a safe bet that we will approach very near a universal solution to the search for solutions. This formulation reminds you of something ? This is normal : it is, in substance, the statement of the Problem of the Price of the Millennium known as the P = NP. In fact, most of the scientists working on the subject agree that it will be difficult to make progress in this direction in the absence of a great revolutionary idea. It is not forbidden to imagine that in the end, this big idea can emerge from an AI self-driven …

Comments