„A neural net chess engine that’s been specifically trained to play like a human of a given level would, by using online games of that level as the corpus“
In the past a „normal engine“ was either unbeatable or played somehow unnatural. This engine could be a gamechanger. What do you think?
Peter
Re: Maia chess - the holy grail for computer chess?
Posted: Sun Dec 06, 2020 4:41 pm
by dkappe
PeterO wrote: ↑Sun Dec 06, 2020 4:16 pm
Is this the holy grail of computer chess?
„A neural net chess engine that’s been specifically trained to play like a human of a given level would, by using online games of that level as the corpus“
In the past a „normal engine“ was either unbeatable or played somehow unnatural. This engine could be a gamechanger. What do you think?
I don't see any type of inspiration citation within the chess engine "makers". The only thanks i saw are for Lichess.
By the way @dkappe, i was wondering if there was a minimum hardware config for these elo-human-nets trained by maia projet to be reliable ? They don't seem to care about that in their paper (or May be I did not find that) .may be because the max elo is 1900 and that can easily achieved by Leela even on a pi 3 cpu ?
Re: Maia chess - the holy grail for computer chess?
reidmcy12/01/2020
Yeah, Elo being a relative measure makes things tricky, we mention briefly in the paper that Leela is a different scale than Lichess. One of our hopes from having bots on Lichess is we'll get actual Elo ratings for our models
[5:57 PM]
dkappe does some great stuff, his work on those small models is what we used as a starting point. I'll test against Bad Gyal when I have a chance. And if anyone else has models we should test against I'd love to hear it
Re: Maia chess - the holy grail for computer chess?
Posted: Mon Dec 07, 2020 4:23 am
by AdminX
Here is an Agadmator vs Maia Video
Re: Maia chess - the holy grail for computer chess?
Posted: Mon Dec 07, 2020 8:07 am
by Madeleine Birchfield
AdminX wrote: ↑Mon Dec 07, 2020 4:23 am
Here is an Agadmator vs Maia Video
Video by ChessNetwork
Re: Maia chess - the holy grail for computer chess?
Posted: Mon Dec 07, 2020 9:55 am
by Madeleine Birchfield
PeterO wrote: ↑Sun Dec 06, 2020 4:16 pm
Is this the holy grail of computer chess?
„A neural net chess engine that’s been specifically trained to play like a human of a given level would, by using online games of that level as the corpus“
In the past a „normal engine“ was either unbeatable or played somehow unnatural. This engine could be a gamechanger. What do you think?
Peter
Cheating online is going to become virtually undetectable. No need for using Stockfish for cheating, only a Maia trained to act like a GM or a strong IM, or for beating up low rated players, a Maia trained on players around 2000 elo.
Re: Maia chess - the holy grail for computer chess?
Posted: Mon Dec 07, 2020 3:44 pm
by PeterO
On their homepage they write: „ In current work, we are pushing the modeling of human play to the next level: can we predict the moves a particular human player would make? It turns out that personalizing Maia gives us our biggest performance gains. We achieve these results by fine-tuning Maia: starting with a base Maia, say Maia 1900, we update the model by continuing training on an individual player’s games. This plot shows that personalized Maias achieve up to 75% accuracy at predicting particular players' moves.“
Image you take a certain chessplayer - his opening book - and 75 % accuracy by his moves!! Fascinating!!
The holy grail?
Peter
Re: Maia chess - the holy grail for computer chess?
„A neural net chess engine that’s been specifically trained to play like a human of a given level would, by using online games of that level as the corpus“
In the past a „normal engine“ was either unbeatable or played somehow unnatural. This engine could be a gamechanger. What do you think?
Peter
Cheating online is going to become virtually undetectable. No need for using Stockfish for cheating, only a Maia trained to act like a GM or a strong IM, or for beating up low rated players, a Maia trained on players around 2000 elo.
Not exactly. The reported accuracy was only a little over 50%.