Almost 25 years ago a machine beat the world’s best at the beautiful, elegant, and brutal game of chess for the first time. We are no match for modern computers, with their ability to analyse every single move combination and outcome with brute force. However, it is only with brute force computing that they can beat us – computers still can’t think and analyse like humans do.
Read: Microsoft profits beat expectations by billions
The best players in the world understand the game so well that they are intuitively aware of only the handful of possible moves that should be considered in any position. A computer needs to analyse every single legal move the find the best. Or at least, they used to. A team of researchers have developed an AI chess engine that tries to play like us, instead of simply crushing us through raw computation power.
Called the Maia engine, it doesn’t necessarily play the best available move every single time. It tried to mimic the way that humans think and has learned an intuitive understanding of the game. This new AI is as a result of a paper written and c-published by researchers from Cornell University, the University of Toronto and Microsoft.
The AI model was trained on individual moves from millions of online human games rather than with the sole aim of winning. This approach allowed the researchers to programme Maia to be able to play at various skill levels. Multiple versions were trained from data of players at various skill levels, tuning the AI to play at a spectrum of skill levels in turn. Nine different AI’s were trained that function between ratings of 1,100 and 1,900 – ratings typical in novice amateurs to strong amateurs.
The AI was first unleashed in December 2020 on the chess site lichess.org. More than 40,000 people played against the AI in the first week, and high levels of engagement has been retained. According to the researchers, Maia matched human moves more than half of the time at each skill level, with accuracy increasing as the skill level increased. Lower rated players make more blunders (big mistakes), so it is more difficult to predict. The system was able to learn what kinds of mistakes players make at different skill levels and recognize the skill level at which people stop making those errors.
