At the end of 2017, DeepMind (British subsidiary of the alphabet dedicated to AI) presented AlphaZero, an artificial intelligence that showed that he could learn to play from scratch, from chess, shogi and Go. and win all the AIs that have been declared champions in each of these games.
Now, in Science magazine, an analysis article about AlphaZero is published. explains how to learn autonomously and refer to a deep neural network through continuous random games and counting more information from the rules of the game.
This meant taking a break from the approach the AI had adopted so far. In the chess field (such as IBM) or from IBM, such as Deep Blue: These were based on thousands of rules and intuitions created by powerful human players trying to explain each event in a game.
Example of reinforcement learning
According to the authors of the study, DeepMind members said, "the results can be learned from zero and a reinforcement and general-purpose learning algorithm To achieve super human performance in various games of great complexity".
The amount of training AlphaZero needed in each case depends on the style and complexity of the game: about 9 hours for chess, 12 hours for shogi and 13 days for Go.
In this case, ir learning through reinforcement an (already used with video games) generates millions of games in a neural network that plays itself in a trial-and-error process. note.
After training, the network is able to focus on the most promising ones based on previous experiences, instead of analyzing all possible movements, using the & # 39; Monte-Carlo search tree & # 39; It is used to guide a search algorithm called.
However (and with less computational capacity), AlphaZero was the winner on all rivals.
An unusual player
DeepMind underscores how they are influenced by members of an innovative, highly dynamic and ğ unconventional ”chess community in the AlphaZero games.
In fact, two international chess players, Matthew Sadler and Natasha Regan, analyzed thousands of AlphaZero chess games for the "Game Changer" book. his style is different from any traditional chess engine:
"It's like discovering the secret notebooks of the great players of the past."
And this, Being taught and not limited to traditional wisdom"By developing their own intuitions and strategies, AlphaZero added a new and broad set of new ideas about the chess strategy that has increased thought for centuries."
For example, if the system calculates that it can benefit in the long term, it is willing to sacrifice parts at the beginning of the game.
"The traditional engines are extremely robust and make a few blatant mistakes, but they can have problems when they encounter positions without a concrete and computable solution," explains Sadler.
"Where AlphaZero claims, you'll see the feeling as & # 39; emotion & # 39; or & # 39; intuition & # 39; In cases that require this is exactly true.".
More from a game
Researchers are proclaiming themselves as "enthusiastic about the creative response of chess," which is a huge challenge for artificial intelligence since AlphaZero's computer era began.
Old world chess champion Garri Kasparov and Artificial Intelligence on '' Deep Thinking & # 39; The author of his book explains the importance of Alpha Zero's achievements in this game: "For over a century, chess has been used as the Rosetta Stone of both human and artificial cognition.".
But AlphaZero team researchers warn that the creation "goes beyond chess, shogi or Go":
"DeepMind's goal is to create a program that can solve some of the most complex problems in the world and teach itself how to deal with chess, shogi, and an important first step on this path." .