A person has long been unable to compete with artificial intelligence in chess. It’s been 15 years since the best human player beat the best AI chess players in a tournament. In general, invincible chess AI has not surprised anyone for a long time. if artificial intelligence learned to play not like a supercomputer, but like an ordinary person. This type of system was developed by a team of scientists from Cornell University, the University of Toronto and Microsoft.
Chess AI Maia
The chess AI named Maia does not calculate its actions a hundred moves ahead and does not necessarily choose the best available move possible. Instead, Maia tries to reproduce what a living person would do.
To create Maia, the researchers used the open source system Leela, which is based on the AlphaZero algorithm from DeepMind. They trained AI on the real moves of millions of live players. At the same time, it was not taken into account whether this move led to victory. The only important thing is that it was made by man.
This approach allowed the researchers to teach Maia several levels of chess skill at once. The AI was trained on the games of users of various skill levels. As a result, nine bots were created to play with people who have ratings from 1100 to 1900 – from a level slightly higher than a beginner to a strong player.
In December, the researchers made Maia available on the chess website. In the first week, people played more than 40,000 games against artificial intelligence. The researchers found that the moves of Maia in most cases corresponded to those that a person would have made. The system was also able to find out what mistakes are most often made by players at different levels and recognize the skill level at which people stop making these mistakes.