Chess Fighter Neural Net release

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
brianr
Posts: 347
Joined: Thu Mar 09, 2006 2:01 pm

Re: Chess Fighter Neural Net release

Post by brianr » Fri Apr 12, 2019 1:40 pm

Alexander Lim wrote:
Thu Apr 11, 2019 7:14 am
Now type "go wtime 5000" and hit enter. If it does a little search then your all ready to install into your chess GUI.
Perhaps you meant go "movetime" 5000" instead of "wtime".

It is playing fine under Winboard.

User avatar
Guenther
Posts: 2935
Joined: Wed Oct 01, 2008 4:33 am
Location: Regensburg, Germany
Full name: Guenther Simon
Contact:

Re: Chess Fighter Neural Net release

Post by Guenther » Fri Apr 12, 2019 2:56 pm

brianr wrote:
Fri Apr 12, 2019 1:40 pm
Alexander Lim wrote:
Thu Apr 11, 2019 7:14 am
Now type "go wtime 5000" and hit enter. If it does a little search then your all ready to install into your chess GUI.
Perhaps you meant go "movetime" 5000" instead of "wtime".

It is playing fine under Winboard.
Here not, at least with HGMs uci adapter, but that seems relevant to my system, it always produces an ucrtbase.dll error
on Win7-64 Ultimate. (all those runtime libraries dependencies are awful...)

With Polyglot it works though, but unlike in CuteChess nothing is recorded in the pgn, no times no evals, probably
because the output is unusual and missing times at all.

(Note that all games were played with a very slow gpu - around 400-500nps in the opening - see pgn header)


Current foe list count : [91 - still rising]
http://rwbc-chess.de/chronology.htm

Alexander Lim
Posts: 32
Joined: Sun Mar 10, 2019 12:16 am
Full name: Alexander Lim

Re: Chess Fighter Neural Net release

Post by Alexander Lim » Sat Apr 13, 2019 3:35 am

Guenther wrote:
Fri Apr 12, 2019 2:56 pm
brianr wrote:
Fri Apr 12, 2019 1:40 pm
Alexander Lim wrote:
Thu Apr 11, 2019 7:14 am
Now type "go wtime 5000" and hit enter. If it does a little search then your all ready to install into your chess GUI.
Perhaps you meant go "movetime" 5000" instead of "wtime".

It is playing fine under Winboard.
Here not, at least with HGMs uci adapter, but that seems relevant to my system, it always produces an ucrtbase.dll error
on Win7-64 Ultimate. (all those runtime libraries dependencies are awful...)

With Polyglot it works though, but unlike in CuteChess nothing is recorded in the pgn, no times no evals, probably
because the output is unusual and missing times at all.

(Note that all games were played with a very slow gpu - around 400-500nps in the opening - see pgn header)
Hi Guenther. Thanks for your time in playing those games. By the way, only two games are on display. The other 10 or so are only viewable if you hit reply. Probably for the best - all loses! :oops:
Hmme not the Chess Fighter I know! When I first discovered that you could do batch inference the nps jumped from around 2k to 11k which resulted in around 300-400 ELO increase. At 400-500 nps I can't imagine what its ELO would be.

Regarding the 400-500nps you reported: Chess Fighter only outputs the total node count so you have to divide by the time taken. Is that what you did?

After seeing those games I felt it my duty defend Chess Fighter's reputation! I managed to download Pawny but couldn't find Abrok 5.05 or Monolith 1.0. It drew with Pawny but then I realised it was Pawny 1.2 (ELO 2778) and not the Pawny 0.3 (2490) that you used. It then also drew with Pawny 0.3. Both games Chess Fighter was winning (according to the engines) but threw away the wins.

I'm currently running a 10 game match against Pawny 0.3. So far +3-0=1 in favour of Chess Fighter. I'll post the games in due course.




User avatar
Guenther
Posts: 2935
Joined: Wed Oct 01, 2008 4:33 am
Location: Regensburg, Germany
Full name: Guenther Simon
Contact:

Re: Chess Fighter Neural Net release

Post by Guenther » Sat Apr 13, 2019 7:31 am

Alexander Lim wrote:
Sat Apr 13, 2019 3:35 am

Hi Guenther. Thanks for your time in playing those games. By the way, only two games are on display. The other 10 or so are only viewable if you hit reply.
You need to click at the 3 dots into the first line of the pgn viewer then you get the whole pgn selection window.
Alexander Lim wrote:
Sat Apr 13, 2019 3:35 am
Probably for the best - all loses! :oops:
Hmme not the Chess Fighter I know! When I first discovered that you could do batch inference the nps jumped from around 2k to 11k which resulted in around 300-400 ELO increase. At 400-500 nps I can't imagine what its ELO would be.
That's why I tried bs20 (batchsize 20) with the cutechess games, but the nps was nearly the same
Alexander Lim wrote:
Sat Apr 13, 2019 3:35 am
Regarding the 400-500nps you reported: Chess Fighter only outputs the total node count so you have to divide by the time taken. Is that what you did?
Yes.
Alexander Lim wrote:
Sat Apr 13, 2019 3:35 am
After seeing those games I felt it my duty defend Chess Fighter's reputation! I managed to download Pawny but couldn't find Abrok 5.05 or Monolith 1.0. It drew with Pawny but then I realised it was Pawny 1.2 (ELO 2778) and not the Pawny 0.3 (2490) that you used. It then also drew with Pawny 0.3. Both games Chess Fighter was winning (according to the engines) but threw away the wins.
from CCRL 40/4:

Code: Select all

Monolith 1.0 64-bit             2819    +15     -15     55.5%   -39.2   34.8%   1471
Abrok 5.0                       2434    +9      -9      44.3%   +41.1   25.8%   5381
Pawny 0.3 64-bit                2463    +13     -13     47.8%   +14.7   27.1%   2206
I wouldn't worry too much about this, the NN size is too big for my gpu and batchsize reduction seems not to help much.

For LC0 with destilled nets or 10b nets I can get over 3200 on this hardware compared to CCRL, but with 20b nets
it will be several hundred of rating points lower. I estimate ChessFighter around 2300 at blitz on my hardware.

It seems ChessFighter is even more dependent on a very strong gpu than LC0 though.
Current foe list count : [91 - still rising]
http://rwbc-chess.de/chronology.htm

Post Reply