Artificial Intelligence sends 5 poker champions to the mat at the same time

Pluribus, its name, managed for the first time to beat 5 players at a time in No-Limit Texas Hold'em. Developed by Facebook and Carnegie-Mellon University in Pittsburgh, Pennsylvania, this artificial intelligence realizes for the first time what no other has accomplished so far. 

In 1997, for the very first time, an artificial intelligence beat a world chess champion ... Since the water sank under the bridges since an AI recently managed to defeat five opponents at Texas Hold 'em, thus making her the first victorious AI in a multiplayer game.

A rather modest configuration

Artificial intelligence Pluribus "formed" a global strategy in 8 days of calculation on a server with 64 cores and will have needed less than 512 GB of RAM. So it is quite small compared to other artificial intelligence deployed in recent years to beat humans in the game. 

Where Pluribus is even stronger, it is that it has not trained against a human, but simply against itself, again and again, to reach its best level.

A very personal way of playing

Alone, for a week, the AI ​​practiced what its programmers have called the Monte Carlo counterfactual regret minimization. The principle of the Monte Carlo is based on the fact of always thinking according to three possibilities and to build three diagrams of the possible games in anticipation of the shots to come, a little like a tree. Regret minimization, meanwhile, is predicting what can be played and putting in place the one that will create the least regrets once the decision is made.

The AI ​​has therefore shown its superiority by avoiding the mistakes that could make other artificial intelligence because, in one against one, it is easy to know the hand of the opponent, but it is complicated when four other people are to take into account. 

Most AIs are thus beaten by finally repeating a habit that the human adversary could spot and put to his advantage. In the same way, Pluribus is able to change tactics in less than a minute, from revival to bluff, making it unreadable and unpredictable. So it's still a victory for artificial intelligence, the question is what will be the next challenge that researchers will be able to meet.