AlphaZero is the name of the new program of the company’s artificial intelligence DeepMind (a division of Google) that, in a matter of hours, became the best player of the history of chess, shogi (a Japanese version of chess) and go. In 1997, when the IBM Deep Blue supercomputer defeated Garry Kasparov chess champion, the society recognized him as a symbol of the progress of artificial intelligence to the human intellect, and thereafter took these challenges as the parameter of the technological evolution. Both DeepBlue and Stockfish (new virtual Chess Champion), Elmo and AlphaGo (the programs developed for shogi and go respectively) use its processing power to analyze millions of moves, items, and sample loaded previously in their systems to choose the best possible option depending on the situation.
The new intelligence of Deep Mind, on the other hand, uses a neural network to only need to know the rules of the game to begin to develop strategies that will enable it to defeat their rivals by itself. It does so through a process of trial and error known as reinforced learning, playing millions of games against itself and adjusting the parameters of the network depending on each victory, defeat and draw. According to the report published on the website Science, AlphaZero took it nine hours to learn chess, twelve hours learning shogi and thirteen days learn go, to then beat Stockfish, Elmo and AlphaGo within four hours, two and thirty hours respectively. To test the true capabilities of the system, the researchers put him in inferior conditions against its rivals, and still went out gracefully all the fighting. Even with a tenth of the time to process the information and confirm a movement, AlphaZero won all the games because their reasoning is more selective: in chess, for example, instead of analyzing the more than 60 million options It includes rival, is limited to a mere 60 thousand possibilities per second since search only among viable movements and with higher chance of success. It is the style of AlphaZero what the chess players found more surprising. The machine arrived on itself to some of the most popular strategies, but not be conditioned by other moves, has developed unique strategies based on his own reading of the game. Consulted on the matter by DeepMind, Garry Kasparov was very surprised the dynamism of the program.
“Instead of processing instructions and human knowledge at a high speed as all previous chess machines, AlphaZero generates its own knowledge,” said Kasparov.
The great master Matthew Sadler stressed “way in which parts surround the King of the opponent with purpose and power” and how the system seems to give less relevance to the “value” of parts. AlphaZero strategies do not prevent you from sacrificing the value in the initial stages of the game if that ensures a profit over the long term.” It is amazing how manages to impose his style of play to a variety of positions and openings,”Sadler said. “Traditional systems tend to be exceptional and make few mistakes, but can show weakness to positions without a concrete and calculable solution. “It is precisely in those positions, in which the ‘intuition’ is required, in which AlphaZero stands.” AlphaZero without a doubt is an amazing program, but several specialists lowering expectations with respect to its application outside of these logical applications. Miguel Lazaro, a Spanish researcher, and Murray Campbell, a specialist in the field who worked in DeepBlue, agree on the fragility of the system, which would require a significant amount of training where the rules change if a little. Lazarus also ensures that this type of disciplines are different “to the stage that usually confronts the human intelligence”, in which we carry out actions which result “only we can foresee partially in environments in which we see a part.”
The victories and defeats of AlphaZero index | “Image: DeepMind Meanwhile Campbell warns about the impossibility of explaining (and hence question) artificial intelligence decisions, something that would make it unworkable in other areas.” While AlphaZero can identify what they believed to be the best movement and provide sequences of movements to back it, is not able to explain their decisions in terms that humans can understand easily,”explained. DeepMind researchers ensure that the ability of AlphaZero to refine three different complex games (and potentially any game of perfect information) is an important step in a clear problem in artificial intelligence : the impossibility of developing skills against any modification of the environment. The goal, ultimately, is to create systems of complex learning that one day may help researchers to find new solutions to some of the most important and complex scientific problems. In this article: GOOGLE DEEP MIND ALPHAZERO ARTIFICIAL intelligence
