AlphaZero, the program that trains itself to play chess (and can beat anyone)

Get improve by playing again and again against itself, without the intervention of their creators The goal is to develop a generic technology of thinking ap

AlphaZero, the program that trains itself to play chess (and can beat anyone)

Get improve by playing again and again against itself, without the intervention of their creators

The goal is to develop a generic technology of thinking applicable in any field

Artificial Intelligence: A new ally of the medical

a Few months before that Carlsen defeated Caruana to win the World Chess championship, took place another duel between two players of a higher level: AlphaZero, a generic program of Artificial Intelligence (AI) fought in a series of 100 games at the best software engine of the world's chess, Stockfish. 28 victories, zero defeats, and 72 tables. With a special feature: the entity that best plays around the world took only nine hours to reach that level, without human intervention, through the reinforcement learning. Now the creators of AlphaZero posted the details of this feat of technology, which occupies the front cover this Thursday in the journal Science.

The scientists of the company DeepMind -a subsidiary of Google and is responsible for AlphaZero - explain that the program not only is able to master the game of chess and other games even more complex (such as Shogi and Go) to levels never seen before, but what you get without having been designed specifically for any of them. And without any prior knowledge beyond the basic rules. "Our dream is to see the same type of system not only in games, but in all kinds of real-world applications; in the design of drugs, materials or biotechnology," explains David Silver, principal investigator of AlphaZero.

it's Been more than twenty years since Deep Blue, the supercomputer of IBM, crew defeated the then world chess champion Garry Kasparov. Although playing chess is not among the most complex tasks that the AI has managed to master, that moment remains a reference point in the take-off of the non-human intelligence. "Shortly after losing my rematch with Deep Blue the small window of competitiveness, man-machine chess is slammed shut forever," writes the own Kasparov in the editorial of Science. "But as demonstrated by AlphaZero, the domain of the machines has not finished with the historical role of chess as a cognitive laboratory".

brute Force vs. learning

in Both Deep Blue as in his successors, the algorithms behind the AI are designed to exploit the features of each game. It is based on a strategic understanding has been worked on by its developers human. In fact, for the development of the computer IBM was fundamental to the participation of great masters, among them the Spanish of Miguel Illescas. By its very nature, chess offers a limited number of moves in each position, which makes possible a approach of brute force; that is to say that the machine you use their processing capacity to dominate the game analyzing all the potential positions.

however AlphaZero represents a new reality, both in the world of technology as in the chess. "Engines and traditional are exceptionally powerful, and commit few errors, but they can get stuck when faced with positions without a concrete solution and measured", explains Matthew Sadler, a Great Teacher and co-author of a book about AlphaZero will be published in January. "It is precisely in these positions, the required sensations, perception, or intuition, in which AlphaZero stands out. Has a style of play very nice."

In this sense from DeepMind explained that the program plays in a much more humane that other forms of IA. Without any policy, tactical, or strategic get better simply by playing again and again against itself, in a method known as 'learning reinforcement'. In reality AlphaZero study, a significantly smaller number of positions than rivals such as Stockfish (about 80,000 per second, in contrast with the hundreds of millions of the engines of chess). Their learning follows the same pattern as that of a human being, only that at a speed completely different. This form identifies the behaviors and patterns more beneficial through trial and error, and then incorporates that information. "Learn directly from the experience, form their own assessments. In that sense, he is free from the limitations in which humans think about the game," says the co-founder of DeepMind, Demis Hassabis.

practical Applications

But, humiliate to humans and to other machines on table games is not the reason why Google invested $ 650 million acquisition of DeepMind in 2014. In reality the goal is to create a technology of thought generic that can be applied to a wide range of challenges. In other words, a catalyst capable of expressing all the potential of the AI. "The goal of DeepMind is to build systems that can solve some of the most complex problems of the real world," says Hassabis.

in Addition, the use of the learning reinforced is not in itself a novelty. But DeepMind you want to combine this method with deep learning (deep learning) to create artificial neural networks of several layers that can be applied to a variety of services, including the famous search engine of its parent company, for example. At present the learning of reinforcement already used in robots laboratory to pick up and move all kinds of objects; it can also detect diseases and ailments in medical scans, or to help tag pictures. And this are only some hints of what is to come.

According to the criteria of

Learn more
Date Of Update: 01 January 2019, 08:00