Defeat of world chess champion Garry Kasparov against Deep Blue in 1997, mastery of go in 2015 and poker in 2019… In the space of a few years, artificial intelligence (AI) has shown that it can surpass top competitors in wildly popular board games.
Alas, the humiliation was not over!
More recently, programs have excelled in even more complex games.
If these feats are more confidential, they are no less impressive for specialists.
Read alsoChatGPT: 5 minutes to understand this artificial intelligence that analyzes and writes (almost) like us
At the beginning of December, the journal Science told how an AI called DeepNash excelled at Stratego until reaching the top of the rankings.
In this game, two players each position 40 pieces on a board, without the opponent being able to see the nature of these pawns.
The goal is to move pieces in turn to eliminate those of the opponent and seize a flag.
Fine strategist
For two weeks in April, DeepNash competed against human players on the online gaming platform Gravon.
After 50 matches, the virtual competitor has achieved the unthinkable, ranking third among all the players who have played on the site since 2002!
Not content with thus reaching an "expert" level among humans, the AI also fined all the robots that paraded on Gravon, including the three-time winner of a world championship between computers.
However, the number of possible courses of the game is dizzying: it is equivalent to 10 power 535, or a 1 followed by 535 zeros!
By way of comparison, the number of potential games in the game of go is “only” 10 to the power of 360. Like poker, Stratego is a game with imperfect information: the hand (for the first) or the positions pieces (for the second) are concealed from the adversary, who therefore advances in the fog.
But if there are only one million possible hidden positions in the famous card game, there are 10 to the power of 66 in Stratego!
To win 84% of the games, DeepNash (developed by DeepMind, owned by Google and famous for having raged in the game of go) proved to be a fine strategist.
As crazy as it may seem, the virtual competitor decided for himself whether or not to take such a piece from the opponent at the risk of revealing the nature of one of his own.
He also showed himself capable of bluffing.
Advance masked
At the end of November, already, it was Meta AI, a laboratory of Facebook's parent company, which struck a blow by announcing that it had developed "the first AI agent to achieve human-level performance in Diplomacy".
In this game, it is no longer two but up to seven players who compete.
Each represents a nation seeking to take control of supply centers on the eve of the First World War.
One of the difficulties is that you have to ally with other players, according to their strengths and weaknesses, and communicate in private.
Subtle maneuvers that Cicero, Meta's AI, performed brilliantly, by aping in the chat module… human players!
“If a computer, which we know is strong by nature strategically, is the only AI at the table, I think the other players, the humans, will rather play against it.
So the AI has every interest in pretending to be a human to blend in with all the players, ”explains Gabriel Lecointre, member of the French-speaking Association of Diplomacy Players (AFJD).
In 40 online games, Cicero "scored more than double the average score of human players and ranked among the top 10% of participants who played more than one game," according to the results of the experiment. also published in Science.
"When AlphaGo beat the best go players in the world, a CNRS researcher said that the next steps would be poker, bridge and Diplomacy", recalls Gabriel Lecointre.
"But no one in the Diplomacy community thought it would happen so quickly..."