DeepMind announces AI that can win the game by negotiating with other players in the diplomatic game 'Diplomacy'
DeepMind, which handles
Negotiation and honesty in artificial intelligence methods for the board game of Diplomacy | Nature Communications
https://doi.org/10.1038/s41467-022-34473-5
AI for the board game Diplomacy
https://www.deepmind.com/blog/ai-for-the-board-game-diplomacy
Diplomacy is a board game set in Europe on the eve of World War I. Each player in charge of seven European powers actively conducts diplomacy with other countries, forms alliances to defeat common enemies, and sometimes allies. It is required to betray and bargain.
For AI, which has won against humans in one-on-one games such as chess and Go, it is necessary to acquire the logical thinking ability to understand complex communication with other players and whether other players keep or break their promises. Getting it on has been a major challenge for many years. Therefore, a research team led by János Kramár, a researcher at DeepMind, first simulated how the game would unfold under various agreements, so that AI agents could select mutually beneficial negotiations. We have introduced an algorithm that allows
But in Diplomacy, promises are not always kept, so agents need to strategically anticipate what will happen if their contracts are broken in order to form alliances with other players. To address this challenge, the research team examined how trust and cooperation are undermined once a promise is broken, and then predicts and predicts retaliation from the broken party. We have trained the AI to only betray if the gain from betrayal is sufficiently greater than the betrayal.
As a result, DeepMind's AI was able to communicate with other agents, make contracts and develop the game advantageously, and was able to win against agents without negotiation ability.
Kramár said of the research, ``Building trustworthy AI with fairness and transparency is a very important theme, so studying these problems in a sandbox like Diplomacy is a real-world challenge. It will help us to better understand the conflict between possible cooperation and competition.' emphasized that the obtained
The ability of AI to make full use of complex strategies is evolving rapidly, and in December 2022, DeepMind's AI ` ` DeepNash '' succeeded in defeating experts in `` Stratego '', a type of military shogi. . In addition, Meta's AI 'Cicero' played Diplomacy online and ranked in the top 10%.
Meta develops AI `` Cicero '' that manipulates `` lies '' to deceive humans and bites into the top 10% of the friendship destruction game `` Diplomacy ''-GIGAZINE
Related Posts:
