An artificial intelligence system called Pluribus, developed by Facebook and researchers from Carnegie Mellon University (United States), has won in poker against some of the strongest Texas hold 'em players, in its six-person version. Over the course of 12 days and 10,000 hands played, the system won an average of $ 5 per hand, totaling around $ 1,000 per hour, with much higher than average winnings. The results of the experiment were published in the scientific journal Science.
First, the researchers taught Pluribus to play poker by making him participate in games in which he had copies of his opponents, following a fairly common practice to make artificial intelligences (AI) learn to perform a certain task. Simplifying, the system learns from its mistakes, playing against itself in thousands of games. Already in this phase Pluribus proved to be very skilled, taking a few days to develop his skills and without the need for particularly powerful computers, when compared with other systems used for the development of artificial intelligences.
Next, the researchers adjusted the AI's behavior so that it worked to predict at most two or three hands following the one in progress, rather than predicting the entire outcome of a game. The study explains that this solution has made possible progress beyond expectations, because in many cases in poker the short-term strategies are more advantageous than the long-term ones, more useful instead for other types of games such as chess or Go.
Pluribus learned to bluff, deceiving professional players who challenged him on several occasions. The AI had no ability to analyze the expressions or posture of the players, details that the pros notice during the games, looking for some clue as to the strategy that others are following. Despite this, Pluribus nevertheless developed the ability to analyze and predict the strategies of his opponents, even when they were bluffing, as one of the researchers explained to The Verge: “AI doesn't see a bluff as something deceptive. It simply sees the decision that will lead to maximizing the payout in a particular situation. What we have shown is that an AI can bluff, and that it can bluff better than a human being “.
Despite the experiment's promising results, it is too early to determine whether an AI is ever more adept at playing Texas hold 'em than humans. It is true that in the 12 days of matches, players have never been able to find any weaknesses in the AI that remained constant, and that could be exploited to their advantage.
Games are often used to experiment with artificial intelligence systems: they have clear and defined rules, more understandable in the automatic learning phase. Three years ago AlphaGO, an artificial intelligence from Google, had beaten the world champion of the complicated board game Go, demonstrating the great potential of the system. The case of Pluribus seems to be even more promising, because poker implies a greater number of variables related to each player's approach to games, and to his choices, which cannot always be predicted with great accuracy.
