Meta’s AI outperforms humans in the classic board game Diplomacy

Researchers at Meta have unveiled the artificial intelligence algorithm Cicero, which plays the board game Diplomacy at a human level. The Register reports.

Diplomacy was designed in 1953. The game focuses on communication and negotiation among players. Each player takes on the role of one of seven European powers at the dawn of the 20th century.

In essence, Cicero is a chat bot that negotiates with other participants to secure effective moves. The algorithm is based on a language model similar to BART, with 2.7 billion parameters.

It was pre-trained on texts from the internet, as well as a dataset of more than 40,000 Diplomacy games played online. The dataset contained more than 12 million messages exchanged by humans during play.

In essence, the AI agent’s dialogue output is anchored to its strategic-thinking module. With it, Cicero constructs ‘intentions’ representing a plausible set of moves by the various players.

“The strategic-thinking module predicts the policies of other players [likely actions] for the current move based on the board state and the overall dialogue, and then selects its own policy for the current move,” the researchers said.

Unlike AI agents for chess, the researchers employed supervised learning using data from past games. However, the approach itself yielded a trusting model that unscrupulous players could easily manipulate, the researchers said.

To prevent such situations, the developers connected Cicero to the piKL iterative planning algorithm. With its help the AI agent refines the initial forecast of another player’s policy and planned moves based on the dialogue between the bot and humans. The algorithm seeks to improve the expected sets of actions by evaluating various options that yield better results.

Over two months, the researchers tested Cicero in 40 anonymous Diplomacy games. According to the trials, the AI agent ranked among the top 10% of participants who had played more than once. In the user ratings for those who played five or more games, the algorithm placed second.

Across all 40 games, Cicero’s average score stood at 25.8%, more than double the average (12.4%) among its 82 opponents.

World Diplomacy champion Andrew Groff praised the algorithm’s dispassionate approach to the game.

“Many people will soften their approach or begin to seek vengeance, but Cicero will never do so. He simply plays the situation as he sees it. Therefore he is ruthless in executing his strategy, but not ruthless in the sense that irks other players,” he said.

According to the developers, Cicero still makes some mistakes. However, they expect their research to be useful for other applications such as long-conversation chatbots or creating realistic video-game characters.

Meta has also released the Cicero source code so that anyone can contribute to its development.

Earlier in November, the tech giant shut down ‘AI for Science’ after three days of its launch. The algorithm was accused of generating fakes and misinformation.

Subscribe to ForkLog’s news on Telegram: ForkLog AI — all the news from the world of AI!