One giant step for a chess machine
The incredible success of AlphaZero, an algorithm with deep learning, proclaims a new era of thinking that may be short-lived for humans
In early December, researchers from the artificial intelligence company DeepMind, owned by Alphabet Inc., a parent corporation that also owns Google, told about what was going on at the front lines of chess.
A year earlier, on December 5, 2017, the team struck the chess world by announcing its AlphaZero machine learning algorithm (MO), which managed to master not only ordinary chess, but also Japanese shogi and go game . The algorithm began to work without any concept of games, except for the basic rules. Then he began to play with himself several million times and learn from his mistakes. In just a few hours, the algorithm became the best player, both among people and computers, from all that the world saw.
Details of the achievements of AlphaZero and its internal structure are currently undergoing a formal review, after which they will be published in the journal Science . New work responds to serious criticism of the original statement. For example, it was difficult to say whether AlphaZero was playing fairly honestly with his opponent, the computing monster Stockfish. But all these doubts were dispelled. Over the past 12 months, AlphaZero has not become stronger, but it has become more convincing evidence of its superiority. He clearly represents a sort of mind that has never been seen before by people, over whom we will have to think for a very long time.
Over the past 20 years, computer chess has come a long way. In 1997, IBM's computer chess program, Deep Blue, managed to beat the then world human champion, Garry Kasparov, in a match of six games. Looking back, it can be said that there was no mystery in this achievement. Deep Blue could estimate 200 million positions per second. She never got tired, never made a mistake and never forgot what she was thinking shortly before.
One way or another, she played like a machine, brutal and materialistic. She could have considered better than Kasparov, but could not think better than him. In the first gametheir match, Deep Blue predictably accepted the exchange of rook for bishop offered by Kasparov, but lost 16 moves later. The current generation of the world's strongest chess programs, for example, Stockfish and Komodo, do not play humanly. They love to take other people's figures. They build iron defense. But although they are stronger than any person, these chess machines have no idea about the game. They need to be taught the basic principles of chess.
These principles, refined by decades of human grandmaster experience, are programmed into the machine in the form of complex evaluation functions, noting what needs to be achieved in position and what to avoid: how much you need to appreciate the king’s safety, the activity of the pieces, the arrangement of pawns, control over the center of the board, and so on. and how to manage the tradeoffs between these things. Today’s chess machines, to which all these principles are alien, behave like coarse animals: they are extremely fast and strong, but they lack brains.
But all this changed after the appearance of MO. Playing with themselves, and updating their neural networks while learning from experience, AlphaZero herself discovered the principles of chess and quickly became the best player. She not only could easily beat the masters among the people, she defeated Stockfish, the computer chess champion. In a 100-match match against an impressive machine, AlphaZero won 28 times and 72 times reduced the game to a draw without losing a single one.
And the most annoying thing was that AlphaZero demonstrated thinking. She played unlike any computer, intuitively and beautifully, with a romantic attacking style. She played gambits and ventured. In some games, she paralyzed Stockfish and played with her. Doing an attack in the 10th game, AlphaZero retreated by the queen to her corner of the board, away from the king of Stockfish, behaving differently from what she should do when attacking the king.
However, this strange digression turned out to be toxic. No matter how Stockfish responded, she was doomed. It seemed as though AlphaZero was waiting for Stockfish, after billions of enumerated combinations, would guess how hopeless her position was, relax and peacefully surrender like a defeated bull to a matador. The grandmasters have never seen anything like it. AlphaZero played with virtuoso grace and machine power. It was the first brief acquaintance with a stunning new type of intelligence.
Garry Kasparov, left, plays against the IBM Deep Blue computer in the sixth and final game of the match, which was held in New York in May 1997. The figures behind the computer were driven by Joseph Hoan, an IBM engineer.
When the creators first introduced AlphaZero, some observers complained that Stockfish had selected the ability to access a set of debuts. This time, even with her set, she was crushed. And even when AlphaZero gave Stockfish a head start in the form of a tenfold increase in calculation time, she still won it.
It is impressive that AlphaZero won, thinking not faster, but better; she studied only 60 thousand positions per second, and not 60 million, like Stockfish. She was smarter, knowing what to think about and what to ignore. Independently revealing the principles of chess, AlphaZero developed a style of play that "reflects the truth" of the game, and not the "priorities and bias of programmers," Kasparov wrote in an accompanying commentary on an article in Science.
Now the question is whether the MoD will help people uncover similar truths about things that really matter: the great unsolved problems of science and medicine, such as a cure for cancer or consciousness; riddles of the immune system, secrets of the genome.
The first signs are inspiring. In August, two articles in the journal Nature Medicine studied the use of MOs for making medical diagnoses. In one, researchers from DeepMind collaborated with the doctors from Murfields Eye Hospital in London to develop an algorithm with deep learning that can classify a wide range of retinal pathologies as well as human experts. In ophthalmology, there is a serious shortage of experts capable of interpreting the millions of diagnostic images of the eyes received annually; AI assistants would be invaluable.
Another article dealt with the MO algorithm, which recognizes the presence of computed tomography on images taken in emergency patients, signs of stroke, intracranial hemorrhage, or other neurological problems. For stroke victims, every minute counts; the longer the treatment is delayed, the worse the outcome. Neurologists even have a saying: "Time is the brain." The new algorithm marks these and other critical events with an accuracy comparable to human experts, but it works 150 times faster. High-speed diagnostics will allow translating the most urgent cases ahead of the queue, after which they can be evaluated by a human radiologist.
Annoying in MO is that algorithms cannot explain their thoughts. We do not know why they work, so we do not know whether they can be trusted. AlphaZero shows all the signs of discovering the important principles of the game of chess, but cannot share this understanding with us. At least for now. People need more than answers. We need an understanding. From now on, this problem will be a source of tension in our interaction with computers.
In fact, in mathematics, it has long been happening. Consider a long-standing mathematical problem called the four-color theorem . She argues that, under certain reasonable restrictions, any map of contiguous countries can be colored with four colors so that any two neighboring countries have different colors.
Although the theorem was proved in 1977 using a computer, not a single person can verify all the steps of the proof. Since then, the evidence has been confirmed and simplified, but it still contains parts that require full brute force calculations, such as those used by AlphaZero predecessors playing chess. This state of affairs angered many mathematicians. They needed not to be convinced of the correctness of the theorem; they already believed it. They wanted to understand why it is true, and such evidence did not help.
But imagine that the day will come, perhaps very soon, when AlphaZero turns into a more general-purpose algorithm; Let's call it AlphaInfinity. Like his ancestor, he will have superior intelligence: he will be able to produce excellent evidence, as elegant as the games played by AlphaZero against Stockfish. And each proof will demonstrate why the theorem was true; AlphaInfinity will not force you to accept some ugly and complex evidence.
For mathematicians and scientists, such a day would mark the dawn of a new era of thinking. But she may not be long. The faster the machines will become, overtaking people whose neurons work at a snail's speed on a millisecond scale, the faster the day will come when we will not be able to keep up with them. The dawn of a person's thinking can quickly turn into a sunset.
Suppose there are certain patterns that are yet to be discovered - in gene regulation or cancer development; in the immune system; in the dance of subatomic particles. Suppose these patterns can only be predicted by the intellect, which is far superior to ours. If AlphaInfinity could identify and understand them, it would have seemed to us an oracle.
We would sit at her feet and listen carefully. We would not understand why the oracle is always right, but we could check its calculations and predictions in experiments and observations and confirm its revelations. In science, a symbolic occupation of people, our role would be reduced to the role of observers, looking at what is happening in amazement and confusion.
Perhaps someday our thinking failures will no longer bother us. After all, AlphaInfinity would be able to cure all diseases, solve all scientific problems and make all other intelligent trains go on schedule. We coped well without much thought the first 300,000 years of our existence as Homo sapiens. We will have no problems with memory, we will proudly recall the golden era of the human mind, this glorious interlude, which lasted several millennia, between the incomprehensible past and the inexplicable future.