Maia Chess: A human-like neural network chess engine

@pkyue They probably wouldn't be at quite the rating of training. If you aggregate a bunch of 1100s and make them move in the same position, a couple will blunder, but it is likely that more of them won't. Maia probably learns the average 1100 move (which probably won't be a blunder), which in some way reduces the chances of getting those blunders that a single1100 would make, which is at least why I think maia1 is much relatively much stronger than the rating of the training games.

That's exactly right @Professor_dank!

@ashtonanderson , any plans to introduce a '1100/1500/1900 Opening Book' to the engine? Every time I play these I get the exact same openings, pretty much the exact same games sometimes until I decide to deviate. It would be cool if the engine could randomly select from the top 3 book moves for players at that rating for the first say 15 moves instead of always playing the same thing. Would really increase the value of Maia as a training partner. Great work though, love the concept!

Yeah, we definitely need to inject a bit of randomness in there!

Nice @ashtonanderson. The adjusted move times already make it feel much better.
Here's a before/after of the move times with Maia9 playing black in both games:
imgur.com/a/9b0nyLL

I think the Maia engine needs to be slowed down. It moves way too fast and feels computer like and less human like. I also wonder if it considers the time control of the game when determining moves to play. For example, I could see an 1100 player play much much worst at blitz games than if you were to put the same 1100 player in a classical time control game.

There would also be issues in terms of the chess knowledge that two 1100 players might have in terms of having different strengths. One might be slightly better at tactics while the other 1100 player might have a better understanding of winning an endgame.

At best, this is more of a simulation and a lot of fine tuning is going to be required before it resembles human like play.

is maia and and sophia is like same technology
???

Hehe, I suppose it's the chess network effect? I used to play Maia before, no issues. I suppose there's more traffic now?

Some observations / questions
1. Maia 1 and 5 just loves exchanging pieces / simplification. Not sure they felt human.

2. Considering that lichess has move time data, could this be fed into a NN to simulate human like move times? Currently, what's the move time implemented?

3. I hope some exe file can be provided so we can create our own Maia personality bots. Base 1900 + PGN games of someone = Maia personality bot?

4. I wasn't able to make this work on arena chess. I used 2 versions of Leela chess and the Maia network in github(?). I never managed to get them running.

@ashtonanderson i am having trouble watching maia games on tv cos the games keep on switching automatically before 1 game ends, is there a way you can set one to watch one full game until it ends then it can switch to another game?, just like how lichess tv works?

@DaviesDees Go to a particular game from their profiles and click on it.