Computers only need 72 hours to play chess to reach the grandmaster level

It has been nearly 20 years since IBM's Deep Blue supercomputer defeated the "best chess player in the world" - grandmaster Gary Kasparov. And since then, "chess machines" have become more and more powerful , so much so that it is difficult to win mobile chess programs when placed at the highest difficulty level.

Computers only need a short period of time to become chess grandmasters

But even though computers are now much faster than they used to be, the algorithms they use remain unchanged for decades. The power of the main machines is based on the ability to search for all possible moves, and gives the most optimal move - a somewhat "sticky meat" method.

And of course, if compared to the ability to calculate walking water, no one can recover with the computer. While the Deep Blue supercomputer is capable of handling up to 200 million chess moves per second, Kasparov grandmaster calculates only 5 per second. However, he can still compete on par with supercomputers, with "personal" skills.

The historic battle between Gary Kasparov and the computer Deep Blue.

Normally when playing chess, we will observe the position of all pieces on the table, then calculate in the moves we feel is most beneficial. This reduces the amount of water needed to count down to several branches, while supercomputers must handle all possible moves.

The consideration of such complicated information is not the strength of computers, but so far that has changed thanks to the contribution of Matthew Lai, a research scientist at the Royal University of London.

He created the machine called Giraffe , with the ability to teach himself to play chess by judging the position of chess pieces on the table , like humans, instead of handling a large amount of water in a way "away meat " like the previous machines.

Although this machine has just been completed, it has reached the level of "the best " chess machine in the world - which has been refined over the years. If you follow the rating of the International Chess Federation (FIDE), the Giraffe machine reaches the level of International Grand Master, equal to about 2.2% of professional players.

The secret behind Lai's system is in neural networks - a method of processing information that simulates the human brain. This system consists of multiple layers of links that will change the method of interconnection depending on the learning process of the computer. Over a period of self-study, this system can produce separate results depending on the information it receives, such as facial recognition in a photo.

Chess - a sport that creates lots of problems for programmers.

In the past few years, with faster and faster computer speeds, it is much easier to refine the network system. Moreover, huge data sets are always available to the computer learning process, making neural networks more powerful than ever. It is these technological advances that help a lot for computer engineers to create larger, more layered networks, even surpass people in areas such as face recognition or writing.

Therefore, it is not surprising that neural network systems can identify different potentials, and this is also the direction that Lai approaches for his system. Giraffe's network of 4 classes together assess the position of the troops on the chessboard in 3 different directions.

The first is to look at the whole of the game, with information like the number of pieces of each side on the table, which is the turn of black or white, can be cast or not, etc. v. . Next is to evaluate each individual chess piece, like where the good army is, where the statue is, . The last step is to mark the cells that the pieces can move to.

Hybrid "train" your system with the data collected in actual games, with possible circumstances. "The teaching of chess systems that each side has 3 Hau children is absolutely unnecessary, because in reality there is no such game," he said. In addition, the system needs to handle even more cases or casualties, even military attachments - which do not occur in professional games. At the same time, the standard data set for training the computer system must be large enough, to cover all the actual situations.

Lai's data set is formed by selecting 5 million any flags in the computer game board. After that, he transformed the chess pieces by adding random but proper moves, before applying this dataset to practice for the computer. As a result, he obtained 175 million different flags.

And the method Giraffe uses to develop himself is to play with himself , with the goal of improving the judgment of the next moves that can take place. This is entirely possible, because there will be fixed times that determine the whole aspect of the match - win, draw or lose. In this way, computers can learn which moves are strong, which are weak.

After Giraffe completes the self-study process, the final step is to evaluate the results. Hybrid testing his system on a basic data set called "Tactical Assessment Set" , including 1500 selected chessstones to evaluate the computer's ability to recognize tactics. "For example, it is possible to use flags to view the ability to handle blank columns, other flags to check how the computer evaluates the value of Code and Statues in different situations, or potential assessment flags. occupying the central part of the board " , he added

Machines now give people very little chance of winning.

The results of this test were rated on a scale of 15000. When the system started its self-training, Giraffe quickly gained a score of 6000, and after 72 hours, the machine achieved a score of 9700 - the equivalent of the best "chess machines" in the world.

"This is amazing, because the evaluation process of the best chess machines today is extremely huge, and refined over the years by both computer engineers and grandmasters. world famous ".

Lai added that this approach has a 46% probability of producing the best move, and assessing the top 3 moves that they consider to be the strongest with an accuracy rate of 70%. So computers can completely ignore all other moves. Of course, this approach is still incomplete, with the weak point being the time it takes for the system to process data 10 times more than conventional chess machines.

However, it does not mean that Giraffe is not able to compete in practice. According to FIDE's evaluation standards, on common computers, this system achieves international Grand Master qualification. Meanwhile, other chess machines are playing at Grandmaster level.

"Unlike other chess machines, the power of Giraffe does not come from the ability to handle a great deal of information in order to have a very far-sighted view of the game. The power of this machine is an evaluation. "It is very difficult to understand the nature of human playing. It is paramount in the beginning and the end, which is the two stages that this machine is the most powerful" - Mr. Lai tich add.

And this is just the starting point. Matthew Lai's next intention is to apply this direction to other games, including chess, the subject in which people are still completely overwhelming chess machines. Who knows, this game will soon change in the future.