As many politicians have argued, diplomacy is an art, requiring not only strategy, but intuition, persuasion, and even trickery, and even the most powerful artificial intelligence (AI) approaches. It’s a human skill that has long been off limits. Meta’s AI algorithm is now showing that it can beat many humans at the board game Diplomacy, which requires both strategic planning and verbal negotiation with other players. According to the researchers, the study could point the way to virtual exercise coaches and dispute mediators. International chatbot diplomacy may not be far behind.
“These are spectacular new results,” says Yoram Bachrach, a computer scientist at DeepMind who worked on the game Diplomacy but was not involved in the new research. “We are particularly excited about Diplomacy because it is a very good environment for researching collaborative AI.”
AI has already beaten humans at strategy games like chess, Go, poker, and video games. dota 2It has also proven to be powerful in natural language processing. generate human-like text continue the conversation with A game of diplomacy requires both. Join his seven players vying for European supremacy. On each turn, players give orders regarding the movement of army and naval units and discuss with other players who can attack or assist. Success usually requires building trust, which is sometimes abused. Both former President John F. Kennedy and former Secretary of State Henry Kissinger were fans of the game.
AI research to date focuses on a version of the game called no-press diplomacy in which players do not communicate. The game’s combination of co-op and competition is itself a challenge for computers, as they have to pursue conflicting goals.New releases released this week chemistry, being the first to achieve respectable results in a complete game. Paper co-author Noam Brown, a computer scientist at Meta, said when he started the project in 2019, he thought it would take 10 years to succeed. “The idea of being able to have an AI that talks strategy with other people, plans things, negotiates things, builds trust seemed like science fiction.”
Meta’s AI agent, CICERO, combines strategic reasoning and dialogue modules. As with any machine learning AI, the module was trained on large datasets. In this case, he’s 125,261 games (both game play and player negotiation transcripts) that humans have played online.
Researchers trained a strategic reasoning module by pitting agents against copies of themselves. It looked a few moves ahead, learning to choose actions based on game state, previous interactions, and predicted actions of other players. During training, the researchers also rewarded human-like play so that the behavior did not confuse other players. It tends to facilitate interaction.
The dialog module also needed tuning. On their own, agents were trained to not only imitate what people would say in the game, but to imitate within the context of the game state, previous dialogue, and what the strategic planning module intended. I have learned to balance In the average game, he sent and received 292 messages that mimicked slang. For example, one message has:Belgium may be targeted, but Den needs your help[mark] next year. ”
Jonathan Gratch, a computer scientist at the University of Southern California who studied negotiation agents and provided early guidance to the Defense Advanced Research Projects Agency program It’s also trying to master diplomacy—noting two innovations. First, CICERO bases its communication on a multi-stage plan, and second, keeps utterances and gameplay within human conventions.
To test its skill, the researchers challenged CICERO to play 40 online games against humans. At least he ranked in the top 10% of players who played two games. Zhou Yu, a computer scientist at Columbia University who studies dialogue systems, said:
Gratch said the work was “impressive” and “important”. But he questions how much CICERO’s dialogue contributed to its success, as opposed to strategic planning. According to the paper, diplomatic experts rated about 10% of CICERO’s messages as inconsistent with its plans or game state. “It suggests it says a lot of shit,” Gratch says. Yu agrees, noting that CICERO sometimes utters non-sequiturs.
Brown says the research could lead to practical applications in niches that currently require human hands. One concrete example: A virtual personal assistant might help consumers negotiate better airline ticket prices. Both Gratch and Yu see an opportunity for agents to persuade people to make healthy choices and be open minded during treatment. Gratch said negotiating agents could help settle disputes between political opponents.
Researchers are also aware of the risks.A similar agent is Manipulate Make political statements, commit financial fraud, or extract confidential information. “The idea of operation It’s not necessarily bad,” Gratch says. “We need guardrails,” such as letting people know they are interacting with AI and that AI doesn’t lie. “Ideally, people are in agreement and there is no deception.”