Code

How neural networks play tricks on people

How neural networks play tricks on people

How to quickly make money in IT in 2025 Try yourself in different areas of IT and take the first step to a new profession remotely!

Learn more

Modern neural networks have reached such a level of development that they are capable of beating humans at most well-known board games. The era of competition between humans and artificial intelligence is over. Now players, scientists, and programmers are working to create effective collaboration with neural networks. This interaction opens new horizons for strategies and approaches in games, allowing participants to improve their skills and tactics. With the help of neural networks, it is possible not only to improve the level of play but also to explore new methods of analysis and decision-making.

To achieve this goal, they had to establish contact with "aliens", teach the computer to make mistakes, and become "centaurs" themselves. In this article, we will describe the process and key moments that led to such extraordinary transformations.

An Alien Game in 24 Hours

Following the AlphaGo program's landmark victory over the outstanding Go player Lee Sedol, developers from DeepMind continued to improve their project. Their goal was to create a universal neural network capable of playing several board games simultaneously. This achievement will be an important step towards developing artificial intelligence that can function similarly to the human brain.

In December 2017, a preprint of a scientific article was published about AlphaZero, a new stage in the development of technologies inspired by AlphaGo. Unlike its predecessor, which specialized exclusively in playing Go, AlphaZero has the ability to learn and improve in various games, such as chess and shogi. This opens new horizons for artificial intelligence and its application to strategy games, demonstrating the versatility and power of machine learning algorithms.

DeepMind specialists conducted an experiment with three copies of AlphaZero, running them on different computers and training them on one of three selected games. The neural network spent 12 hours mastering shogi and only 9 hours learning chess. Ultimately, the program needed less than a day to learn three complex games. This result underlines the efficiency and speed of training of modern neural networks in mastering strategy games.

The neural network was trained using reinforcement learning (RL). The network independently played over 700,000 games against itself, receiving rewards for actions that led to victories. This process allowed AlphaZero to learn from its own mistakes and adjust the parameters of the neural network, which contributed to more accurate decision-making in future games. This training approach significantly improved the neural network's performance and enabled it to achieve outstanding results in a variety of games.

At the beginning of AlphaZero's development, programmers provided the system with only the basic rules of the game. The program began training from scratch, without any prior knowledge or experience, and without having seen a single human game. It received no information about the traditional strategies and methods used by grandmasters at the beginning, progress, and conclusion of chess games. Thanks to this approach, AlphaZero was able to independently develop unique strategies and tactics, which became the basis for its successful performance against the strongest chess engines.

AlphaZero has developed unique and highly effective strategies that deviate from traditional methods of play accepted by chess players. It often employs unexpected moves, such as sacrificing a bishop or moving the queen to a corner of the board, which at first glance may seem counterintuitive. These actions are aimed at achieving a positional advantage, allowing AlphaZero to create winning positions that are difficult to predict.

Peter Nielsen, second to world chess champion Magnus Carlsen, compared AlphaZero's actions to the play of aliens. He noted that AlphaZero is an artificial intelligence that has long competed with other programs but has no experience playing against humans. This comparison highlights AlphaZero's unique approach to chess and its unusual strategies that may be unfamiliar to experienced chess players.

I've always wondered what would happen if aliens débarquer on Earth and demonstrate their Chess skills. Now I have an idea of ​​what this might look like.

Peter Heine Nielsen, second to renowned chess player Magnus Carlsen, occupies an important place in the chess world. His experience and strategic approach to the game help Carlsen achieve high results on the international stage. Nielsen not only supports his protégé during tournaments but also makes valuable contributions to match preparation, developing tactics and analyzing opponents. Chess is not only an individual competition, but also a team effort, where the role of a second plays a key role in a player's success.

DeepMind co-founder and talented chess player Demis Hassabis also agreed with this assessment, describing AlphaZero's playing style as "alien."

AlphaZero demonstrates a unique style of play that differs from traditional human and software approaches. This artificial intelligence employs something unusual, almost alien, that makes its play of chess seem like a process from another dimension. This approach to chess opens new horizons in understanding strategy and tactics, highlighting the capabilities of modern artificial intelligence technologies.

Demis Hassabis, the founder of DeepMind, is a key figure in the field of artificial intelligence. His work and achievements in this field generate great interest among both specialists and the general public. DeepMind, founded in 2010 and acquired by Google in 2015, has made significant contributions to the development of machine learning algorithms and neural networks. The company's research covers various areas, including gaming, healthcare, and process optimization. Hassabis actively promotes ideas about how artificial intelligence can improve people's lives and solve complex problems. MIT Technology Review highlights the importance of his work and its impact on the future of technology.

To evaluate AlphaZero's performance, the DeepMind team organized a tournament in which the neural network competed against computer programs previously recognized as the best in their category. The results of this competition demonstrated AlphaZero's clear superiority over all competitors, confirming its outstanding abilities in artificial intelligence and strategic thinking.

In chess, AlphaZero competed against the Stockfish program, in shogi, it faced the Elmo algorithm, and in Go, it faced AlphaGo, its predecessor, which defeated Lee Sedol in 2016. These confrontations demonstrate AlphaZero's outstanding abilities in artificial intelligence and machine learning, as well as its potential in complex strategic games.

In each of the matches, 100 games were played. AlphaZero convincingly defeated all of its opponents in chess, without losing a single game. In shogi, it lost only eight games to Elmo. However, in Go, the results were less clear-cut, though AlphaZero's victory was still significant.

Results of the game AlphaZero against the strongest computer programs Infographics: Skillbox Media

During the game, AlphaZero analyzes a significantly smaller number of possible moves compared to programs of previous generations. For example, in chess, AlphaZero processes approximately 80,000 positions per second, while its main competitor, Stockfish, evaluates approximately 70 million variations in the same time frame. This indicates that AlphaZero uses more efficient algorithms and deep neural networks to evaluate positions, allowing it to achieve high results despite the smaller volume of moves it processes. This approach demonstrates a fundamentally new level of machine learning and artificial intelligence in the field of chess programs.

AlphaZero demonstrates high performance because it focuses on analyzing the most promising strategies. This is achieved through a deep understanding of the game situation, which allows the algorithm to accurately determine key directions for further action. Some researchers call this ability of neural networks "digital intuition," emphasizing their ability to make informed decisions based on the analysis of large volumes of data.

A year after the launch of AlphaZero, specialists from DeepMind presented a detailed description of this program in an article published in the journal Science. However, AlphaZero's source code was never published, and the program was not made available to the general public.

In 2018, a group of enthusiasts created an alternative to AlphaZero, based on the original program's operating principles described in scientific papers. They developed a free, open-source chess engine called Leela Chess Zero. This project marked a significant step in the development of chess technology, offering players access to a powerful tool for analysis and play. Leela Chess Zero uses deep learning and neural networks, allowing it to improve its skills and adapt to different playing styles.

Like AlphaZero, Leela Chess Zero is built exclusively with the rules of chess, and the neural network initially has no pre-training. Leela Chess Zero is trained on a distributed computing network, which is managed through the project's website. In its first few months of operation, Leela played over 23 million games against itself, achieving grandmaster status and continuing to improve. This makes Leela Chess Zero an outstanding example of artificial intelligence in chess, capable of continually evolving and improving its game.

In January 2022, the project team presented a report on the achievements of its neural network, which had played over 500 million chess games. It was also noted that Leela Chess Zero plays approximately one million games daily. These results underscore the system's high efficiency and popularity in the chess world.

Based on the Leela Chess Zero source code, which is available on GitHub, programs have been developed that grandmasters use to improve their skills and prepare for chess tournaments. These tools allow you to analyze games, study strategies, and improve your game, making them indispensable in chess practice.

You have to be able to make mistakes, too

Playing with a superhuman artificial intelligence that thinks outside the box is certainly exciting, but it doesn't always bring practical benefits. In preparation for important chess matches, grandmasters will be more effective using programs that imitate the playing style of people rather than alien beings. The use of such programs allows for a better understanding of the strategy and tactics of human opponents, which in turn helps improve the skill level and prepare chess players for decisive matches.

A team of American scientists led by Professor John Kleinberg of Cornell University has developed a project aimed at making the playing style of a neural network more "human." This project is called Maia, in honor of the outstanding Soviet chess player and world champion Maia Chiburdanidze. The goal of Maia is to improve the interaction between artificial intelligence and players, and to increase interest in chess through more natural play. Building on Leela Chess Zero, the developers changed the approach to training the neural network for chess. Instead of using reinforcement learning, in which the program learned by playing against itself, they implemented a supervised learning method. This allowed the neural network to analyze and study numerous games played by people online. This approach significantly increased the effectiveness of training and improved the quality of the game.

During the training process in Maia, the position of the pieces on the board is given, taken from one of the games played by people. A neural network must learn to predict what move a chess player would make in this situation. Infographic: Maia Chess

The programmers used recordings of grandmaster games from the well-known chess website Lichess as a training base for their algorithms. Lichess is a platform where professionals from all over the world can find suitable sparring partners and improve their skills.

Research has shown that "humanity" in chess largely depends on the types of mistakes players make. Beginner chess players are more likely to make simple miscalculations, known as "blunders," than more experienced players. Therefore, the task of training the Maia artificial intelligence is to enable it to predict and reproduce mistakes common to chess players of different skill levels. This will help better understand human play and improve the quality of interaction between players and programs.

Maia project specialists organized a training database of chess games, categorizing it by the players' ELO ratings. This resulted in nine sets of training games, grouped by ELO rating ranges in 100-point increments, starting from 1100 and ending at 1900. This allows chess players to more effectively study games and analyze strategies depending on their opponents' skill level.

Each training set used 12 million games, with an additional 500,000 records allocated for testing the network's performance after training. This resulted in nine neural networks, each demonstrating abilities similar to human play corresponding to one of the nine ELO rating levels. These neural networks provide high accuracy and realism in reproducing gaming strategies, making them a valuable tool for analyzing and improving gaming skills.

In 2020, the Maia project team published the source code on GitHub. They also presented the results of their work in a scientific paper titled "Fusing Superhuman AI with Human Behavior" and on the project's official website. This publication marked a significant step in the field of artificial intelligence, showcasing the team's achievements and opening new horizons for research in this field.

According to recent research, Maia demonstrates the ability to predict player actions in online games with 52% accuracy. This may seem insignificant, but it's important to understand that it's not a simple "hit/miss" outcome. This level of prediction opens new horizons for artificial intelligence development, enabling the creation of more adaptive and intelligent systems capable of learning from user actions and improving the gaming experience.

The Maia neural network demonstrates impressive performance in chess, as a chess player typically considers approximately 35 possible moves at any given moment. Previously, existing programs could only predict players' actions on the chessboard 35-40% of the time, significantly inferior to Maia's results. This system uses advanced machine learning algorithms for analysis and decision-making, making it one of the most advanced programs in the field of chess.

Maia has a unique ability to predict not only successful moves but also blunders in a game. It can identify even the most unexpected blunders and awkward moves that chess players sometimes make. This remarkable ability to anticipate mistakes makes Maia a valuable tool for teaching and training players, helping them avoid common mistakes and improve their skills. Using Maia in training promotes a deeper understanding of the game and increases your chances of success in chess games.

Maia can analyze your games and identify which mistakes were predictable and which were random. This will help you pinpoint areas for improvement and focus on those aspects of the game that need to be worked on to reach the next level of skill.

In this example, Maia neural networks with an Elo level of 1100-1400 predict that White will make an enticing but strategic The incorrect move b6. In contrast, Maia versions with an Elo level of 1500–1900 predict that in the same situation, players with a rating of 1500 and above will make the tricky move bxa6, forcing the opponent to expose the queenside. Infographic: Maia Chess

The creators of Maia set themselves the task of further training the neural network so that it could imitate the playing style of famous grandmasters. To achieve this goal, they used a neural network pre-trained on millions of chess games, feeding it a limited number of games played by a particular master. This allowed Maia not only to improve its skills but also to adapt its strategy to the characteristics of the selected player. This approach ensures a more accurate reproduction of playing styles and increases the level of engagement with users looking to learn and develop their chess skills.

The results of the study showed impressive results. After analyzing 10,000 games played by a renowned chess player, the program was able to predict his future moves with an accuracy of 65% to 75%. This confirms the effectiveness of algorithms in chess analysis and move prediction, opening new horizons for the use of technology in intellectual games. Maia can be further trained on games played by chess players such as Magnus Carlsen, Anatoly Karpov, or Mikhail Tal, allowing the creation of an ideal sparring partner. This system simulates the play and mistakes of a chosen grandmaster with a high degree of accuracy. Practice against virtual opponents based on the best chess minds of the past and present offers unique opportunities to improve players' skills. The opportunity to compete with the legends of the chess world makes the training process more exciting and effective.

Imagine that you are preparing for the decisive match for the title of World Chess Champion against Magnus Carlsen. You have a program that can simulate Magnus's playing style. This is a unique opportunity to analyze the strategies and techniques used by one of the greatest chess players in history. The program will help you study his tactics, anticipate moves, and prepare for high-level matches. Using this technology can significantly improve your chances of success in the fight for the championship title.

Matthew Sadler is a British grandmaster and the author of "Game Changer." His contributions to the world of chess and literature on strategy are of great interest to chess players and amateurs. "Game Changer" illuminates modern approaches to the game, helps players improve their skills, and provides new ideas for analyzing games. Sadler shares his experience and knowledge, making his work valuable to a wide audience.

On lichess.org, you can play against neural chess programs with Elo levels of 1100 (bot maia1), 1500 (bot maia5), and 1900 (bot maia9). The platform also features a video of a chess blogger from the US testing Maia.

The author was impressed by the skill and "human" playstyle of the neural network, despite initial skepticism. In the conclusion of the video, he noted: "Maia is actually a really impressive technology!"

User reviews of the game against Maia are available on reddit.com. Here you will find a variety of opinions and discussions that will help you better understand the features and impressions of the game. Check out the comments of players and find out what they think of their gaming experiences.

Impressive work! I played a few games against a level 1100 bot and I must say that its play was really similar to a human one. It successfully passed my Turing test, indicating a high level of artificial intelligence.

The program's concept is truly appealing. It is especially useful for people who experience anxiety or nervousness when playing against strangers. This program offers unique opportunities to enhance the gaming experience and reduce stress, making it an excellent tool for anyone who wants to enjoy the game without unnecessary anxiety.

The Maia developers are confident that the results of their research can be applied not only to chess but also to other fields. One example is the prediction and prevention of errors in X-ray analysis. Such technologies can significantly improve the accuracy of diagnostics and improve the quality of medical care.

One of the effective solutions is the use of our developments in the field of diagnostics based on the analysis of medical images. This program is designed to support doctors in the diagnostic process, allowing you to identify images with a high probability of medical error. This tool can significantly improve the accuracy of diagnostics and reduce the risk of errors, which in turn improves the quality of medical care.

John Kleinberg, a professor at Cornell University, is a prominent scientist and the developer of the Maia system. His research in computer science and data processing has made significant contributions to the development of technology. Kleinberg actively works on topics related to algorithms, data analysis, and their application in various fields. His ideas and developments continue to influence modern methods of working with information.

Centaurs Enter the Game

Grandmasters, observing the success of computer programs, have begun to consider the possibility of turning these technologies into their assistants, not their opponents. The combination of strategic thinking, characteristic of the human mind, with the analytical superiority of computers opens up new prospects in the world of chess. This collaboration can lead to a deeper understanding of the game and improved skills for both professional and amateur players. The integration of computer technology into training and match analysis can significantly improve the level of play, allowing chess players to use powerful algorithms to develop and test new strategies. Experts call a system consisting of human-computer interaction a "digital centaur." When properly organized, such a pair demonstrates significantly higher results than either participant alone—both human and computer program. Effective collaboration between humans and technology opens up new opportunities for increased productivity and improved work quality in various fields.

Human reasoning plays a key role in controlling a chess game at a high level of abstraction, forming overall strategy and tactical plans for the game. At the same time, neural networks and other computer models are capable of performing complex calculations, providing hints, and minimizing typical human errors. By harnessing the power of artificial intelligence, players can improve their skills and make more informed decisions, making the game more engaging and strategically rich.

The first chess match using computer hints took place in 1998 between Veselin Topalov and Garry Kasparov, assisted by ChessBase 7.0 and Fritz 5. This match marked the beginning of a new era in chess, leading to the concept of "advanced chess," where players can turn to artificial intelligence for assistance. This innovation changed the approach to the game, opening new horizons for chess players and creating a unique symbiosis of man and machine in intellectual sports.

In recent years, advanced chess tournaments have been held, which have contributed to the refinement of the rules and mechanisms for interaction between humans and artificial intelligence. These competitions have become a platform for developing strategies that effectively combine human intellect and AI computing power in the context of a digital centaur.

In 2007, a significant chess event for Russian chess enthusiasts took place in the GUM exhibition hall in Moscow. The advanced chess match pitted World Champion Viswanathan Anand against former champion Vladimir Kramnik. Watching advanced chess is significantly more engaging than traditional chess matches, as it utilizes modern technology and computers, adding an element of strategy and innovation to the classic game.

Chess competition organizers often display images from grandmasters' monitors, which are connected to chess computers, on a large screen. Spectators have the opportunity to observe in real time how the players interact with artificial intelligence, calculating various game continuations and choosing the best ones. This provides the audience with a unique opportunity to see the thought processes of leading chess players in action, significantly enriching the experience of chess matches and making them more accessible to the general public.

Digital centaurs are finding applications beyond chess. In 2021, Russian programmers presented the world's first effective "human + AI" system for playing Go. As part of the 41st World Go Championship, digital centaurs played 70 matches, competing both against each other and against pure artificial intelligence. This development demonstrates new possibilities for human-AI interaction in strategy games, opening up prospects for further technological development and its application in various fields.

Schoolchildren and students who won the "Brain Games" hackathon developed computer systems for collaborative work between humans and artificial intelligence. Eight development teams arrived in Vladivostok to create unique versions of human-machine interfaces for Go. These innovative solutions demonstrate how technology can unite humans and AI in intellectual games.

The games were played on a smaller version of the board measuring 13 × 13 cells, as opposed to the standard board of 19 × 19.

At the championship At the World Go Championships in Vladivostok, centaur players could contact a neural network for help. Photo: Center for Regional and International Cooperation.

Players could interact with the artificial intelligence through a special interface, receiving computer-generated hints. However, for each hint used, the centaur's points were deducted as a penalty. This mechanic added an element of strategy to the game, forcing players to weigh the need for a hint against the potential loss of points.

The best Russian players, Ilya Shikshin (7th dan EGF and 4th professional dan) and Alexander Dinerstein (7th and 3rd dan, respectively), faced off in virtual battles. The players used an interface developed by the "Black and White" team from Bauman Moscow State Technical University. In the "digital centaur" format, they competed against the Leela Zero artificial intelligence, based on the neural network, and emerged victorious. These tournaments highlight the skill level of Russian Go players and demonstrate the development of artificial intelligence technologies.

According to tournament organizer and Go grandmaster Oleg Mezhov, the Leela Zero artificial intelligence demonstrates outstanding results on a 13x13 board, leaving even the most experienced players virtually no chance of victory. This ability of AI underlines its influence on modern strategy games and opens new horizons in the study of Go.

Defeating artificial intelligence at Go is almost impossible, but today a significant event occurred. Teams of humans and artificial intelligence, playing on the same side—the so-called "digital centaurs"—were able to defeat artificial intelligence for the first time. This event demonstrated that the combination of human intuition and machine power can create a unique force capable of overcoming modern technology.

Oleg Mezhov, the Russian and European Go champion, is the organizer of the tournament in Vladivostok. His achievements in the world of Go inspire many fans of this intellectual game. As a leading player, Mezhov actively promotes Go in Russia and abroad, organizing tournaments and events that draw attention to this ancient sport. The Vladivostok tournament he hosts is becoming a milestone in the Go community, providing players with the opportunity to showcase their skills and share experiences.

The collaboration of neural networks and humans in games such as Go and chess serves as a shining example of how artificial intelligence can effectively work together to solve complex problems. This collaboration demonstrates how AI can complement human intelligence, elevating the game and opening new horizons in strategy and analysis. The application of neural networks in these gaming disciplines not only improves results but also provides the opportunity to study complex algorithms, which could be useful in other fields such as business, medicine, and science.

Systems similar to chess "centaurs" could significantly improve the effectiveness of interactions between natural and artificial intelligence in various areas of life. However, humans will remain key—they will continue to make decisions and bear responsibility for them, even when relying on recommendations provided by neural networks. This collaboration between humans and machines opens up new horizons for optimizing processes and improving the quality of decisions.

Read also:

  • AI cheating in iconic board games: how neural networks learned to beat people using intuition
  • Volleyball, radio engineering, and data analytics: how to become an IT specialist after graduating from a technical university
  • Launch arguments in C#: what they are and how to use them