For years, humans have struggled to compete with artificial intelligence in chess. It's now been 15 years since a human player defeated a computer in a tournament. However, researchers have developed an AI chess engine designed to mimic human play rather than dominate it.
The Maia engine, co-created by scientists from Cornell University, the University of Toronto, and Microsoft, takes a unique approach. Instead of always opting for the best possible move, Maia aims to reflect the decisions a human player would make. This innovative AI was built using the open-source Leela platform, which is inspired by DeepMind's AlphaZero.
The researchers trained Maia using data from millions of online chess games, focusing on individual moves rather than winning strategies. This method allowed them to tailor Maia to different skill levels by creating multiple versions of the AI that cater to players with ratings ranging from 1100 to 1900—covering novice to strong amateur levels.
In December, Maia was launched on lichess.org, quickly amassing over 40,000 games played within the first week. The results were promising: Maia matched human moves more than half the time across all skill levels, with accuracy improving as player skill increased. Researchers noted that Maia could learn the common mistakes made by players at various levels, identifying when players stop making those errors.