Maia, not Maya

In my previous TCEC/CCC interstitial post, A Mind-boggling Tactical Battle (November 2020), I wrote,

No one can accuse me of overusing an easy idea. It's been ten months since I last featured a video on this weekly engine series.

Now you can start accusing. The video on this current post was on my short list for this month's featured video, The Chess Boom of 50 Years Ago, and was too good to pass up. Why too good? Because it led me on a new path of discovery.

Maia Chess: A human-like neural network chess engine (34:48) • '[Published on] Dec 1, 2020'

I've already mentioned ChessNetwork (aka Jerry), the maker of the video, several times on this blog, as the search box in the right navigation column will verify. The video's description starts,

Microsoft researchers and collaborators at the University of Toronto and Cornell University have created chess AI that better matches human play at various skill levels. Unlike AlphaZero and LeelaZero which learned through self-play, Maia learns from millions of online human games.

At this time nine versions of Maia have been trained, one for each Elo milestone between 1100 and 1900. Maia 1100 for example was only trained on games between 1100-rated players, and so on. [...]

The video has already attracted more than 35.000 views and more than 360 comments. As usual, right-click the embedded video to find its address on Youtube. Here are a couple of important links included in its description:-

• Maia Chess (maiachess.com) • 'A human-like neural network chess engine : Maia’s goal is to play the human move -- not necessarily the best move. As a result, Maia has a more human-like style than previous engines, matching moves played by human players in online games over 50% of the time.'
  
  
• The human side of AI for chess (microsoft.com) • 'Chess stands as a model system for studying how people can collaborate with AI, or learn from AI, just as chess has served as a leading indicator of many central questions in AI throughout the field’s history.'

When AlphaZero was announced three years ago, its developers explained that it couldn't train on human games, because there were too many errors in the play. It turns out that there's much to be learned from that style of play as well.