Introducing Maia, a human-like neural network chess engine

Can we play with Maia????

@ISHADEEP click on the crossed-swords icon on her page to set up a game. She has three versions:

Easy : @maia1
Medium : @maia5
Hard : @maia9

She seems to only accept live games with a time limit, not correspondence games.

@ggCaden exactly, correspondence games give the bots troubles and moves they make don't depend on the time control so hopefully classical is long enough for most people

@rfw Cheaters are something we've been worried about too, one of the longer term goals of the project is to better understand what it means to make human moves in chess so we think the project will make cheating harder in the long run. Also, our chess engines are deterministic (they use an opening book in the early game) so it would be relatively easy to check if someone is playing with one. Plus there's not that many people that want to cheat so they can play like a 1400 rated player.

@reidmcy But what if your team wanted to develop Maia bot was of players around 2200-2300 rating (or a bot that specifically played like say, Magnus Carlsen). This could lead to rampant cheating which could be not easy to detect.Also, a 900 rated player would not mind using a 1400 rated Maia bot..

I played 4 games with Maia9, the final score was 3-1(in favor of me), she definitely makes human blunders but 1900 rating is a bit unstable I would say like in 1 game, she blundered the queen with a basic pattern like queen on d4 and my bishop on d3 and then I sacced it on h7 to win her queen, in one game, she beautifully outplayed me on dark squares, showing human like intuition. It would be cool to see how this project turns out, I have one doubt, it seems as if she is getting better as we play more games (if that's the case pls tell me)

@The_Human_Paradox Those are very valid concerns. Our group as talked about the issue of impersonation and don't plan on releasing a model like that. Also, we have done some experiments with copying highly skilled players and I don't think our models can imitate very strong players they play too much like Stockfish.

@sam27 We're not that advanced the bots are the same as what's in github.com/CSSLab/maia-chess every time. We just use an opening book for the first few moves so the bots don't always play the same game

Amazing work, Lichess team!

Please make a stronger Maia or work to tighten her blunders. I screwed up in the opening against Maia9 and around halfway through she just seemed to let me win. Is anyone else experiencing this?

Like this, overwhelming loss turned into repetition draw:

Or this, early blunder turned into a win & drawish position turned into easy win:

For a "1900" she plays more like a 1600, where solid play can often recover from a blunder if you just keep some pressure on them. FYI, I'm a 1750 so I should not win 2 out of 3 against a 1900.

@reidmcy I Played a few games in bullet against maya5, during endgame with little time remaining it definetly doesn't play like a human, all its move are almost premove even with pieces disadvantage. This has to be corrected...

However, except for time management, it's awesome, very human like!!!

As for making a stronger Maia, couldn't you restrict her learning algorithm to only learn from players above x rating. That's not to say she wouldn't play weaker opponents, that's what can make her an awesome practice partner, but she'd ignore the data if they weren't good enough.

i.e. Maia9 could only learn from players rated between 1800-2000 and whenever she plays players between 2000-2200 it could help create Maia11. Once you have Maia 11, it could learn from players rated 2200+ to eventually create Maia13. Also, if you need more data, couldn't you just feed her every game in your database between players of the appropriate skill level?

My Python skills are only lower-intermediate and my AI skills are very weak so I could be way off, but it seems like it'd cut down on her quality volatility.