By the end of 2017, DeepMind (British subsidiary company assigned to the AI) AlphaZero, gave false information that showed that it was possible to play learning from starting chess, shogi and Go , The and ending its & # 39; winning the AI named as campaigners in each game.
Now, the Science magazine has published an analysis article about AlphaZero explains how he started self-learning and progressing to a weird deep network, through regular random games and counting more information before the regulations; game.
That meant breach of the approach taken to date by the AI which is excellent in chess field (such as Stockfish, or Deep Blue from IBM): they were based on thousands of rules and heuristics that were created by strong human players trying to explain every event in a game.
Example of consolidation learning
According to the authors of the research, DeepMind members, the results show that reinforcements and algorithms can learn a common purpose from learning. start and achieving superhuman performance in a number of games of great problem"
The level of training required by AlphaZero in each case was dependent on the style and complexity of its game: about 9 hours for chess, 12 hours for shooting and 13 days for Go.
The statement of learning is strengthened & # 39; (already used with video games) and including this issue in a strange network that will be able to access the video. Play millions of games against yourself in a trial and error process, in a way giving awareness of the type of plays that you have; adds to the goal that has been winning its & # 39; game.
Once it is trained, the network will be used to guide a search algorithm that is & # 39; tree-search Monte-Carlo lets you, instead of analyzing all possible moves, AlphaZero can be at # 39; focus solely on the most gelator according to their previous knowledge.
Despite this (and have a lower computing ability), AlphaZero was named as winner of all competitors.
DeepMind describes how they responded to chess community members, which they saw in the AlphaZero games of innovative style, a very lively and unusual game.
Indeed, two international chess players, Matthew Sadler and Natasha Regan, made thousands of thousands of AlphaZero chess games for their "Game Changer" book (which will be released within a month) and says the style is different from any traditional chess engine:
"It's like to find out the secret notebooks that have been a good player of the time."
And that is, Being self-taught and not limited by ordinary wisdom, AlphaZero created his own ideas and strategies, contributing to a new set of new ideas that he has; Increase the centuries of thinking about chess strategy. "
The system, for example, is particularly willing to give parts at the start of a game if it works that it can get a long-term benefit.
"Traditional engines are very strong and they will not make clear mistakes, but they can have difficulties when dealing with a non-concrete and responsive situation," a & # 39; Sadler explains.
"It is straightforward in those situations, in which feelings & emotions, & # 39; ideas & # 39; not & # 39; essential, where AlphaZero says"
More on a game
The researchers say they are "enthusiastic about AlphaZero's creative response to chess, which has been a major challenge for artificial information from the beginning of computing age."
Garri Kaspárov, a talented chess campaigner and author of the book & Deep Thinking & # 39; on uniforming information, explaining the importance of # 39; What is Alpha Zero's performance in this board game: "For more than a century, chess is used as a Rosetta Stone of human and voluntary knowledge"
But AlphaZero team researchers have warned that their creation is "passing over chess, shogi no Go":
"DeepMind's goal is to build systems that can solve some of the most difficult problems in the world and create a program that can be learned by learning how to continue chess, shogi and & # 39; starting from the first important step on that route ".