Meta's AI learns how to play Diplomacy
While Meta's (the parent company to Facebook) primary pursuit has been expanding into the world of virtual reality, it appears they're also teaching AIs how to play board games. AI is already beating humans at the classic games Go, chess, poker, and even DOTA 2. And recent advances allow natural-language processing to generate humanlike text and contextual conversations. The game of Diplomacy has 7 players try to take control of Europe, and most definitely relies on these skills. Each turn, players issue orders to move army and naval units, and follow discussions with the other players, while attempting to build trust (and sometimes abuse it) as they gather support (only to stab them in the back later).
Former President John F. Kennedy and former Secretary of State Henry Kissinger were both fans of Diplomacy. Noam Brown, a computer scientist at Meta, co-authored the paper detailing the project which started in 2019. He originally thought it would take a decade to complete.
"The idea that you can have an AI that’s talking strategy with another person and planning things out and negotiating and building trust seemed like science fiction." -Noam Brown
The AI agent, named CICERO, uses strategic reasoning and dialogue modules trained on large data sets. In this instance those data sets are 125,261 games that humans played online along with their transcripts of player interaction and game plays. It further improved itself by having the AI play games against copies of itself to learn actions based on the state of the game, previous dialogue, and the predicted actions of other players while learning the balance between deception and honesty. It even mimicked game slang to appear more human. CICERO played 40 online games against humans and placed in the top 10% of players who’d played at least two games.
The question to seasoned Diplomacy players is … will it be as satisfying stabbing an AI in the back as it would be a human player?