Meta’s new AI plays a famous board game as well as a human


CICERO can play diplomacy as a human and talk to his opponents to devise complex strategies.

Hearing him talk about metaverses indiscriminately would almost make you forget that Meta is also a leading player in the world of artificial intelligence. Mark Zuckerberg’s troops yesterday unveiled their latest advancements in this area; they presented CICERO, an AI that delivers impressive performance on Diplomacy, a famous board strategy game.

At first glance you may wonder what is so special about these works. After all, artificial intelligences have proven time and time again that they can humiliate people in very complex games. Gary Kasparov, a chess legend who reigned supreme over the discipline for years, can attest to this; he bowed to DeepBlue way back in 1997 (see our file on the history of computer chess).

And since then, these algorithms have continued to evolve. In 2016, AlphaGo defeated Go god Lee Sedol. This was a remarkable new breakthrough, because for years this game, which relies heavily on intuition, was considered unaffordable for these types of algorithms. Rebelote in 2019, when an AI beat a professional poker table for the first time. We can also mention many video games such as DOTA 2 and Starcraft IIwhere man no longer has the slightest chance of defeating a specialized program.

Algorithms that confront the human psyche

But all these games have a common denominator: the interactions between players are completely based on game elements such as cards, pawns or characters which can be represented by a handful of variables. For example, it is possible to transcribe an entire game of chess using a simple sequence of letters representing the pieces, and the coordinates of the squares they move to.

In the case of diplomacy, however, the situation is quite different. The action is certainly materialized by pawns, but is essentially based on the negotiation and verbal interactions between players. However, these exchanges imply a certain capacity toabstraction Additionally; Algorithmically, this is very different from chess, where every move (such as “Queen in h4”) can be easily translated into computer language.

Here we must take into account the subtleties of the language, but also the way each player interprets the situation. To win a game of diplomacy, the AI ​​must be able to forge alliances, evade bluff attempts, gauge how each human will react to their blows… and for an AI that is much more difficult than calculating the optimal option from a well -defined list of possible moves. However, Meta claims to have succeeded with CICERO.

CICERO, a formidable strategist

This program operates through two separate subsystems. The first analyzes the forces present to propose a strategy, while the second takes care of the exchange with the players and the interpretation of their comments. This is the first time that the same program has been able to ‘discuss’ textually as well as construct new strategic reasoning based on the responses it collects. And apparently the recipe works; according to Meta, the AI ​​invariably scores twice as high as its flesh-and-blood opponents.

The program even managed to place itself in the top 10% of players on, the online version of the game, making it a formidable opponent from a strictly strategic point of view; scholars now consider it achieved” performance on a human level “.

© © MetaAI

To take the analysis further, Meta’s troops called in a specialist; like DeepMind, which asked European champion Fan Hui to improve its AlphaGo program, they recruited World Diplomacy Champion Andrew Gof.

His expertise helped to highlight some significant shortcomings, and unsurprisingly, this mostly pertains to the human and emotional component of the game.” Many human players can soften their approach or be driven by a sense of revenge, but CICERO never does. says Goff.

This tends to make the program more predictable than a human in certain situations, because it is systematically preferable the approach that seems to him mathematically the most optimalwithout necessarily considering the long-term plans of his opponents.

Learn to communicate with AIs

But what makes these works interesting is that their scope goes far beyond diplomacy. According to Meta engineers, this technology could be very useful in many other contexts as well.

Mastering natural language generation can facilitate communication between humans and AI-based systems “, they explain in their press release. For example, such an algorithm could make it possible to develop educational AIs capable of ” have a long conversation with the aim of learning new skills to a human.

We can also imagine numerous applications related to entertainment, such as in the case of video games. “Imagine a video game where an NPC can plan and converse like a human — understanding your motivations and customizing the conversation — to help you on your quest,” the post’s authors suggest.

What is surprising, on the other hand, is that not once does the press release mention the famous metaverse that Mark Zuckerberg dreams of. And knowing Meta’s roadmap in recent months, it would be very surprising if the company did not try to implement this kind of technology in this famous virtual space. It will therefore be interesting to see how and in what context the technologies underlying CICERO will be reused in the future.

Similar Posts:


Please enter your comment!
Please enter your name here