It was in 1997 that IBM Deep Blue won a historic win over world chess champion Gary Kasparov. Now, in the world of chess, it’s time to say that it is almost impossible for a living player to beat the highest level AI.
However, AI has different characteristics from humans, and in the past, some players took off and won. Cornell University, University of Toronto, and Microsoft joint research team unveiled Maia, an AI chess engine that learns a lot of human thinking and thinks like a human, rather than pointing to the best hand with vast precalculation results.
Maia developed based on Leera Zero, an open source AI chess engine created based on a 2017 paper on Alpha Zero, Google’s deep mind AI. The research team trained AI by dividing it into nine grades using millions of notation data for human players, ranging from beginners to semi-pros online. So Maia can fight against all players with moderate strength.
In December 2020, the team unveiled Maia on an online chess site (lichess.org), with more than 40,000 matches in the first week. According to reports, Maia fought in more than half of the battles with a whistle, according to each ranking. Not only did the whole thing be done without errors, it became possible to play the game like a human being.
The research team said that the current general chess AI does not have the concept of making certain mistakes by humans at a certain level, so it can point out the mistakes made by the matched player, but cannot show how to solve them. The AI chess engine has shown the potential to become a learning tool, saying that the algorithm is characterized by what typically occurs and needs improvement in
In addition to chess, AI has surpassed human abilities in various fields. Now, the research team is looking for ways in which AI, which has only been self-improving until now, can learn something jointly with humans. For example, in the medical field, diagnosis results can be presented through CT scan image interpretation. However, algorithms couldn’t directly teach humans something because the way they approach the problem was completely different from that of humans.
This study is not a good learning method in terms of AI learning efficiency, but it does show the potential for AI to take more human-like actions than it is now. Related information can be found here .