Artificial intelligence has surpassed humans in teamwork ability

2018-01-20 07:15

Artificial intelligence has surpassed humans in teamwork ability
The creators of the work decided to check how successful would be the computer in cooperative games.

Sign up for news “UkrMedia” in Facebook, Twitter or Google+

An international group of researchers tested the algorithm, training the computers to find mutually beneficial solutions, reports Rus.Media. When testing the effectiveness of the technique with people it became clear that the two computers are able to agree among themselves more efficiently than two people. A study published in the journal Nature Communications, says Naked Science. The developers believe that their study will eventually help in creating artificial intelligence, with well-developed communication skills with people.

Artificial intelligence becomes the rival of man in games where there are winners and losers: in chess, checkers or go.

The creators of the work decided to check how successful would be the computer in cooperative games. They players join forces to achieve the best outcome for all team members.

The researchers tested 25 available algorithms that analyze the progress and results of games played parties.

In test games involved teams from two computers or two persons, or human-computer. In the first phase of the work none of the algorithms failed to find a successful long-term strategy game, best for both players in the team. In the second phase, the researchers added to the conditions of the experiment the ability to exchange small messages. According to game theory, people establish a cooperative relationship, including using “empty talk” (cheap talk). This is the type of communication that does not require much effort, but can indirectly affect the course of the game.

If successful, the developments of the computers were sending messages like “Great! We’ll be rich!”, when you try to violate the agreement — “You betrayed me!”.

Using these messages, the algorithm S# learned to adjust the game and cooperate with the partner, ensuring mutual benefit. The rest of the game teams of the two computers took mutually beneficial solutions in 100% of cases, while the human players only in the 60%. The computers learned to use the “empty talk” is so natural that people who are in the same team with the computer, could not with certainty determine who their teammate — or algorithm.

The researchers believe that their work will help to use the mathematical framework of cooperation for the development of “social” artificial intelligence, able to cooperate.