By 'superhuman' computing power, the new AI system defeats world poker players

Computer scientists have successfully developed a new card game system, called Pluribus. Why can Pluribus be confirmed successfully? It could beat some of the world's leading poker holdings, hold a number, become one of the most important breakthroughs in artificial intelligence.

Two years ago, the research team at Carnegie Mellon University developed another poker system called Libratus , able to bring down the world's top poker players in stressful 1v1 displays. Libratus' father, Tuomas Sandholm and Noam Brown have overcome their own limits, successfully developing systems that can play Texas hold'em poker in a 6-person table. The result this time is a combination of AI lab of Facebook and Carnegie Mellon University.

During the experiment, Pluribus demonstrated his ability to defeat his opponent, with the performance rated as 'superhuman'. When it came to the 'make money poker' screen , Pluribus won convincingly and earned an average of 1,000 USD / hour.

Pluribus played convincing winning poker and earned an average of $ 1,000 per hour.

For decades, AI researchers have been successful in training artificial intelligence to play 'simple' games: all information is in front of them, only facing a single opponent. most, every move exists for both players to see; The two most obvious examples are chess or Go.

Poker is more difficult in that most of the information about the game is hidden - like cards on the opponent's hand or still in the set, the player's decisions in a poker game make the game even more part. difficult to guess and complicated. The more players you bring in, the more difficulty will be added.

For AI researchers, poker is the game worth the practice system. Because life is not like chess board, with all information lying on the table and clearly identifying the winner. By giving artificial intelligence to handle hidden information to make the best decision, computer scientists can extend the scope of AI application to many other areas.

For AI researchers, poker is the game worth the practice system.

' We do not focus on any industry, just think that this study can apply to many aspects such as network security, phishing detection, enemy tactical analysis, even assisting cars. Drive on crowded roads , 'Brown said.

In the new study, Sandhold and Brown let Pluribus face two tough tests. The first test required Pluribus to face 13 professional poker players, one of whom also earned more than $ 1 million in his career. A poker game will consist of 6 players, a Pluribus facing 5 other professional players.

In the second test, Pluribus faced two world poker legends, Darren Elia and Chris 'Jesus' Ferguson. Each of them will be against 5 Pluribus machines in a 6-card game. If Pluribus had a sweat gland, it would probably have bathed in the face of Darren Elia - who won four World Poker Tour titles, currently holding a record number of awards and Chris 'Jesus' Ferguson, who won 6 World Series Poker events held in Las Vegas.

In the first test, 5 players and Pluribus had a total of 10,000 cards and played all 12 days. A total of 14 people (only 13 people and one device, in fact) will be divided into US $ 50,000, so no one will know whether they are playing with a machine, nor the identity of their opponents.

In the second experiment, Elia and Ferguson faced 5 Pluribus machines in turn. They took up a total of 5,000 items.

Every victory of Pluribus has been confirmed 'beyond all statistics'.

In all cases, all of Pluribus's victories are confirmed 'beyond all statistics'. In the research report, they call Pluribus's power superhuman - superhuman .

' We use that word because machine performance is higher than the best brains ,' Brown said. ' Even when playing with the player, every 100 times, the machine won about 5 times the big blind, the players rated this as a very high winning rate '.

In simple words, if converting this winning rate to real money, then Pluribus will hold about 1,000 USD / hour. Again, this is Pluribus facing poker players! Computer scientist Roman Yampolskiy, who did not appear in the new study, agreed that Pluribus had supernatural capabilities.

'If the machine can show supernatural power by defeating the world's top players, it can defeat the inferior people, meaning that its ability in this area belongs to the super class. ' Mr. Yampolskiy said.

This is a remarkable achievement, because unlike a Go or a chess game, poker contains a lot of hidden information and has good luck, which means that a machine will not be able to surpass the player just by ability. Simple rate calculation. From the early days of humanity's development of artificial intelligence, poker has been the target.

From the early days of humanity's development of artificial intelligence, poker has been the target.

Before the actual test, Pluribus was trained by playing poker with himself. After 8 days of continuous fan spread , it produced its own strategy.

'Pluribus does not use poker player data to create tactics on its own. It does so by playing itself, up to trillions of cards to produce a basic strategy. The more you play, the better the strategy and the better the results, " explained Brown.

'With Pluribus, we make a new strategy of finding tactics, without asking the machine to find out how to win the whole poker game. Instead, it can stop after playing a few games. The breaks make algorithms can scale up more efficiently. Specifically, it allowed us to achieve superhuman computational power with short training time, costing less than US $ 150 to hire cloud computing services to train Pluribus, it was still playing over time. real with only two CPUs'.

Even with a short training time, Pluribus still surpassed the world poker player.

Most importantly, Pluribus is programmed to become unpredictable - the ability is crucial for a person to play poker successfully. If Pluribus constantly pushes big money when he knows the hand has good cards, the opponent will recognize it immediately. To overcome this, Pluribus is trained to play in a balanced way, not in any money-oriented way, preventing players from guessing Pluribus's intentions.

Pluribus is trained to play in a balanced way, not in the direction of any money.

There are two things Pluribus surprised the poker players:

1. The ability to donk bet is very effective. This is a deposit step that usually only entrants to poker, or must be carefully calculated to donk bet at the right time. Even high-ranking players rarely use this method, but Pluribus proves to show them that if they have the superhuman computational power of a computer, donk bet turns into an effective 'money-making' tool.
2. The ability to place huge amounts of money, which poker players often avoid. This tactic allows Pluribus to push the player into a difficult position, when he knows how high his hand is.

" Once again, artificial intelligence surpasses people without data on how to play from humans ," said researcher Yampolskiy. 'It means that machines can teach themselves how to solve complex problems without human intervention '.

Yampolskiy was not surprised by the power of Pluribus, and wanted to see how it managed on a 10-person poker table, not bound by the rules (in the new experiment, Pluribus was not allowed to push in more than $ 10,000).

The power of calculation and situational analysis of AI is progressing. One day, more powerful machines than Pluribus will make decisions that affect the economy, politics so reasonably that we cannot fail to listen.

Just hope the machine is not smart enough to be aware, forever the senseless machine.

New research has been published in Science.