Maia Chess: A human-like neural network chess engine

very nice bot.
the bot tried really hard in the endgame when being down 2 pieces, most humans are tilted when down a lot of material and play worse.

but how can i play

Its opening is pretty good. But its middlegame and endgame is pretty bad. It is basically like a human who plays opening like a grandmaster and middlegame like an unrated player. this feedback is for maia 9. I shouldn't be able to beat a 1900 fide rated player so easily.

It also says gg in the middle of the game. " On a scale from 1 to 5 how humanly did I play? 1 = not human, 5 = very human"
1

pretty bad bot

#33 I partially agree. the bot misses some very simple tactics that any intermediate/advanced player would see in a moment. i have played it only in classical time control, so in blitz things can be different, i don't know.

in this game i blundered badly two times, and still managed to win after the bot blundered a queen fork

i feel like the strength of the bot is measured by the average centipawn loss it sustains... so after a pretty equalish and boring game it has to blunder badly just to reach 50/70 centipawn loss ratio. i can be wrong, but it feels very unrealistic. even a weak player like me can play pretty accurate games from time to time

anyway, i really like this bot, i hope you will be able to improve its calculation, add an opening tree and make it able to play different first moves, and add some other stronger bots... i'm really looking forward to play them for training!!! very very interesting bot. keep up the good work, gl

(thank you Jerry, i LOVE your channel)

my attempt to play is always declined?

imo, the maia1 and maia5 are far stronger than an 1100 and 1500 lichess player, respectively. However, maia9 seems quite a good bit weaker in the couple of games I've played against it (or maybe it's appropriately tuned and the other two are far overtuned). This probably stems from the fact that an aggregation of many 1100 or 1500 players (which, correct me if I'm wrong, is a simple way of modeling Maia) is much less unlikely to make the one/two move blunders characteristic of players at that level than a single 1100 or 1500 player would. However, by 1900 there are much fewer simple blunders from the individual, which means a less noticeable difference between Maia9 and an actual 1900 in absolute terms.

Another point of contention I have with it is its openings. It seems to always play the same variations, which, while maybe technically more "human", is less useful for training purposes. Also, I think it moves a tad too fast when its time is critical. While I really like that it speeds up towards the end, I'm not sure how realistic it is that a human, particularly one below 2000, would move instantly. Overall, I love the concept though

I like this mate that maia9 allowed, I've definitely fallen for stuff like that before. h6 is a very human move

I'm not a very aggressive player, but I've noticed maia1 falls for backrank mate easily. As others have said I've noticed her openings are pretty similar. I'm going to start using the black pieces against her, I observed her smash a poor guy with the london system, running up the H file with rook/queen. I had mistaken which side was who until black resigned lol.....so at least that game she appeared very human like. More fun than stockfish AI for now, but the repetitive openings could become boring. (yeah, yeah, I missed mate in 1 for 2 moves until I saw it.......)

Oh, waouw, I'm really extatic about this project! Thanks a lot ashtonanderson and all your team for bringing this marvel to us. So to the question: does it play like a human? Well, for the overwhelming part, yes it does and that's absolutely amazing! Some things I've noticed to enable maia to play even more human-like: 1) As pointed out by chessnetwork, I feel like maia's time management is a bit too optimal: I'd love (especially in fast control games) to be able to flag it in a very dirty way (mwahahaha). You replied that you had that on the radar, so I'm eagerly looking forward to it. 2) Maia doesn't learn at all from its mistakes. Now this wouldn't be a problem if you play only one game (or two if you switch colours) but as a human, although I have my set openings, if I get crushed in one game in a specific line, I try to deviate at some point if I play the same opponent again in the same line. It may seem very hard to implement, but if you somehow manage to make it play differently depending on how you played vs it beforehand (and without improving its level too drastically at the same time which sounds complicated) it would be absolutely amazing! 3) Maia "only plays chess": it never surrenders (which I believe is easily circumvented) and never ask for a draw. It would be a superb improvement if Maia could ask for a draw (even sometimes when it wrongly assumes it stands worse) or in very marginal cases surrender in non-losing position (as it happened to the best of us... sigh ^^). 4) Now, this one will sound weird but maia plays almost "too" humanly to be completely human. It plays almost only natural moves, sometimes do silly mistakes as we all do (and that point is actually a very good one since the silly mistakes all looked natural or at least explainable to me) often walk into tactical traps (in my opinion just a tad more than it should but at least that's not an unpleasant sight xd) and never goes crazy for sacrifice sake whether it's dubious or sound. Humans on the contrary sometimes go berserk, they are creative and do sacrifice from time to time even when sometimes it is unsound. Overall, I think it underscores that maia's playstyle is a bit too positional to me at the time being. Humans have a wide variety of playstyles but I think most people do not play as positionaly oriented as Maia. Another unrealistic idea that I randomly throw: to create several playstyle versions of maia to give it different playstyles. As for the inherent level, I think it'll settle by itself with more and more games played. I read many people think maia is underrated, while to me it looks a bit overrated, but since I'm the only one thinking that, I'm probably wrong. Sorry for having written "A La Recherche du temps perdu" but I'm so enthusiast about this project that I needed you to have as much feedback as I could give! Continue the excellent work!