The developers trained artificial intelligence to make mutually beneficial decisions in cooperative games using "empty talk" - the exchange of simple phrases-reactions. At the same time, teams consisting of two computers turned out to be more honest players than people.
An international team of researchers tested an algorithm that trains computers to find mutually beneficial solutions. When testing the effectiveness of the technique with the participation of people, it turned out that two computers are able to communicate with each other more effectively than two people. The developers believe that their research in the future will help in the creation of artificial intelligence with developed skills for communicating with people.
Artificial intelligence regularly becomes a human rival in games where there is a winner and a loser: in chess, checkers or go. The authors of the new work decided to test how successful the computer will be in cooperative games. In them, the players join forces to achieve the best result for all team members.
Researchers tested 25 existing algorithms that analyze moves and results in games played. The test games involved teams of two computers, two people, or a person and a computer. At the first stage of the work, none of the algorithms was able to find a successful long-term strategy of the game, beneficial for both players in the team. In the second step, the researchers added the ability to exchange small messages to the experimental conditions. According to game theory, people establish cooperative relationships, including through cheap talk. This is a type of communication that does not require much effort, but can indirectly affect the course of the game. With a successful development of events, computers sent messages like “Great! We will get rich! ", While trying to break the agreement -" You betrayed me!"
With the help of such messages, the S # algorithm has learned to adjust the course of the game and cooperate with a partner, seeking mutual benefit. By the end of the game, teams of two computers made mutually beneficial decisions 100% of the time, and human players only 60%. At the same time, computers learned to use "empty talk" so naturally that people who fell into the same command with a computer could not confidently determine who their teammate was - an algorithm or a person.
The researchers believe their work will help leverage the mathematical underpinnings of cooperation to develop cooperative "social" AI.
The research is published in the journal Nature Communications.
Natalia Pelezneva
Promotional video: