Meta (formerly Facebook) created an artificial intelligence (AI) capable of beating human opponents in games of the online game Diplomacy. It is a popular strategy game in which seven players compete for control of Europe by moving pawns around a map. Unlike other AI-powered board games like chess or go, Diplomacy requires players to talk to each other – to form alliances, negotiate strategies – and know how to recognize when others are bluffing.
This AI, called Cicero, ranked among the top 10% of players who finished more than one game. In total, Cicero played 40 games and faced 82 human opponents who didn’t know they were dealing with a bot. After an eight-game tournament involving 21 players, Cicero took first place. Meta described his work in an article published on November 22, 2022 in the journal Science.
Learning to play Diplomacy is not easy for several reasons. Not only does the game involve multiple players performing moves at the same time, each round is preceded by a brief negotiation where players chat in pairs with the aim of forming alliances or taking on rivals. After this round of negotiation, the players then decide which pawns they will move and whether or not they should honor their pact.
Throughout each moment of the game, Cicero has modeled how other players are likely to act based on the position of pawns on the board and his conversations with them. The AI then determines how players can work together for their common benefit and generates messages intended to achieve those goals.
DeepMind’s playful AI broke a 50-year-old record
Meta’s AI made some mistakes
To develop Cicero, Meta combines two different types of AI: a reinforcement learning model that determines which moves to perform and an extended language model that negotiates with other players.
Cicero is not perfect. He always sent messages containing errors, sometimes contradicting his own plans or making strategic errors. However, Meta claims that Cicero’s human opponents have often chosen to collaborate with him over other players.
And that’s an important thing to consider because while games like chess or go have a winner and a loser, real-world issues don’t usually have such simple outcomes. Finding compromises and workarounds is often more important than winning. Meta says Cicera is a step towards AI that can help solve a range of complex problems that require trade-offs, from planning routes to bypass heavy traffic to negotiating contracts, for example.
Article by Will Douglas Heaven, translated from English by Kozi Pastakia.
This blockchain video game lays the foundation for a metaverse that no one could control
Receive our latest news
Every morning, the information to remember on the financial markets.