Wednesday , January 27 2021

Scientists develop an AI program that teaches itself



The researchers presented AlphaZero, a computer program capable of not only getting superhuman masters in some of the most complex board games known as chess, shogi and go, but also teaching themselves to play them without prior knowledge, apart from the rules of each game . "

The system was created by experts from DeepMind, a British company owned by Alphabet, the parent company of Google.

The results, researchers said, represent an important step in the development of an AI player, who can quickly learn and manage any game.

Chess, Shogi and Go

Since the IBM Deep Blue chess program defeated Gary Kasparov in 1997, the game AIs has progressed and is able to defeat people into more complex games.

Even in other abstract strategies, such as Shogi and Go, both significantly more complex than chess, they were also dominated by machines.

Fast self-learning

AlphaZero's news, developed by David Silver and his colleagues in DeepMind, is that it avoids the need for information derived from humans and, instead, is able to learn chess, shogi and go playing each other until they manage each one of the games

The system has managed to defeat the most advanced AI programs specialized in these three games after only a few hours of self-learning.

As a result, researchers say, they must now turn to a new generation of games, such as multiplayer games, to deal with the next series of challenges for AI systems.

<! – – Download a joint document to this news ->


Source link