After chess and before our delicious brains, AI comes to social strategy games – and it wins | So Good News
Ever since the chess computer Deep Blue defeated world chess champion Garry Kasparov in 1997, artificial intelligence has increasingly dominated humans in several “adversarial” games in which human interaction plays a limited role. Now, however, a team of researchers has discovered a new AI that tries to expand the pool of games a computer can beat you at.
In it paper (opens in new tab) Published this week, researchers report that Cicero, an AI trained to win Diplomacy Games, is a seven-player board game in which “every turn involves all players. […] “free dialogue with others during conversations”. This discussion phase distinguishes Cicero’s actions from other AIs.
According to the paper, “almost all previous advances” in AI have been in “two-player zero-sum” games, in which gaining an advantage for oneself puts the other player at a direct disadvantage. In these games – Chess, StarCraft, Go and Poker – AI can learn optimal strategy by playing against itself in a pattern known as “self-play”. Eventually, he will come up with a way that he can’t win in a balanced game. In these examples, the difficulty of the game is not important; the important thing is that communication is not a central game mechanic, and every action tends to bring the other player back to their goal.
This is inconsistent with Diplomacy, in which communication between players is important (if not exhaustive) and winning does not necessarily hurt the opponent. Here, self-play “created an uninviting language.” This was a major hurdle to overcome, as anonymity was key to a fair experiment. Communication between the players had to be based on the state of the game, or what had already happened, and if Cicero slipped up, he could be caught because he couldn’t explain his mistake.
More important, however, was the ability to build trust with other players. In theory, this concept would be foreign to Cicero, but success requires establishing “the ability to think about the beliefs, goals, and intentions of others” as well as “the ability to persuade and build relationships through dialogue.”
To install Cicero, the researchers took a dataset of more than 40,000 dialogue-based Diplomacy games from the online version of the game. A basic dialogue model was trained on Diplomacy chat logs and then trained to predict messages based on an array of game data. Finally, Cicero was trained to “use” the information in the message to decide his next course of action, while also considering what the other players might try to do.
Finally, Cicero entered the online league anonymously, which runs from August to October 2022. He played in 40 games, ranking in the top 10% of all-time starters and second among the 19 players who played more. five games. Overall, Cicero won the tournament, averaging twice as much as some of his 82 competitors.
It couldn’t be completely eliminated, but this tournament-winning effort for AI laid a solid foundation for future efforts like this one. For now, it may be limited to Diplomacy, but it strikes me that Cicero-like technology could one day make its way into games like Settlers of Catan, or even social deduction video games like Salem Town or Among Us. Right now that’s it would be a sauce.
Need some experience before new AI bosses? These are the ones The best board games outside.