Page 8 of 9
Re: Train your engine !
Posted: Fri Mar 03, 2023 7:51 pm
by deeds
Here is a summary of the "learning session 9" openings :
You will find individual learning curves
here.
learning session 9 :
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%) W D L D(%) OppAvg OppN
1 brainlearn 19.2 nn-1e7ca356472e : 1 4 7269.5 14500 50.1 3009 8521 2970 58.8 0 1
2 Eman experience : 0 ---- 29577.5 58000 51.0 12282 34591 11127 59.6 -7 4
3 ShashChess 25.3 nn-1e7ca356472e : -6 4 7129.0 14500 49.2 2841 8576 3083 59.1 0 1
4 HypnoS 051122 nn-1e7ca356472e : -8 4 7087.5 14500 48.9 2749 8677 3074 59.8 0 1
5 Aurora 301022 nn-1e7ca356472e : -16 4 6936.5 14500 47.8 2528 8817 3155 60.8 0 1
White advantage = 72.64 +/- 1.00
Draw rate (equal opponents) = 50.00 % +/- 0.00
Re: Train your engine !
Posted: Fri May 05, 2023 7:18 pm
by rulleeeee
Hi. When I run pgnToScores.exe it shuts down immediately. Maybe I'm missing something? I have Windows 11
Re: Train your engine !
Posted: Fri May 05, 2023 7:29 pm
by deeds
rulleeeee wrote: ↑Fri May 05, 2023 7:18 pm
Hi. When I run pgnToScores.exe it shuts down immediately. Maybe I'm missing something? I have Windows 11
Not sure what is the main problem (user rights, w11 incompatibility, missing file, etc.) but you have the source code here :
https://github.com/chris13300/pgnToScor ... modMain.vb
This tool contains less than 120 lines so it's not an heavy app.
You should ask people for help there :
https://outskirts.altervista.org/forum/ ... =41&t=3567
Re: Train your engine !
Posted: Fri May 05, 2023 7:32 pm
by rulleeeee
I know absolutely nothing about programming, but the fact that I downloaded one file that was there. WHAT else do I need?
Re: Train your engine !
Posted: Fri May 05, 2023 8:56 pm
by deeds
rulleeeee wrote: ↑Fri May 05, 2023 7:32 pm
I know absolutely nothing about programming, but the fact that I downloaded one file that was there. WHAT else do I need?
The command line of this tool is :
pgnToScores.exe path_to_your_pgn_file.pgn
But you can also do a drag'n'drop of a pgn file onto the exe file.
ps : you have a PM at infinitychess forum
Re: Train your engine !
Posted: Thu Jun 01, 2023 6:29 am
by deeds
Re: Train your engine !
Posted: Thu Jun 01, 2023 1:40 pm
by deeds
Here is a summary of the "learning session 10" openings :
You will find individual learning curves
here.
learning session 10 :
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%) W D L D(%) OppAvg OppN
1 brainlearn 19.2 nn-52471d67216a : 0 4 8004.0 16000 50.0 148 15712 140 98.2 0 1
2 Eman experience : 0 ---- 32232.0 64000 50.4 739 62986 275 98.4 -3 4
3 HypnoS 051122 nn-52471d67216a : -3 4 7936.5 16000 49.6 75 15723 202 98.3 0 1
4 Aurora 301022 nn-52471d67216a : -3 4 7925.0 16000 49.5 27 15796 177 98.7 0 1
5 ShashChess 25.3 nn-52471d67216a : -4 4 7902.5 16000 49.4 25 15755 220 98.5 0 1
White advantage = 4.92 +/- 0.96
Draw rate (equal opponents) = 50.00 % +/- 0.00
Re: Train your engine !
Posted: Fri Jun 02, 2023 10:07 am
by deeds
Hi folk !
It seems incredible but in 2023, so after months/years of
explanations,
evidence,
experiments, there are still daily testers who organize tournaments with engines set to training mode !?
In training mode (=default settings), the engines try some moves without knowing their efficiency like they did when they are learning an opening.
Main problem : these engines are playing in a tournament ! Normally, they are supposed to use the most effective moves they have learned.
Second problem : these testers publish the results of their tournaments as if they were representative of the experience data.
But since we know these testers will never figure out how to set up their engines in tournament mode (=played moves with a conformity with experience > 95%), we keep several endless sources of laughs.
If we use engine's releases newer than the Nov. 5, 2022
patch, all our databases and experience files become obsolete because the scores scale has changed.
Instead of an evaluation based on a pawn value, the new scores scale is a WDL model based on fishtest games, so on selfplay games, so it has nothing to do with the pawn value anymore.
Furthermore, this WDL model changes several times per year, it is no longer suitable for long-term training.
Don't update an old experience file with these new scores else you will corrupt it.
So you have to restart from scratch !
Re: Train your engine !
Posted: Fri Jun 02, 2023 10:07 am
by deeds
As in the past few months/years, I already did ten Eman's learning sessions (of which I have revealed to the public the
results), I'm not going to restart from scratch.
So i only use engine's releases using the old score scale.
As these engine releases are no longer compatible with the new sfnnv6 nets
i will avoid publishing the results of my next learning sessions in a general public area.
Out of respect for the enlightened, i will publish them on my personal
website.
I thank the SF team for Stockfish and their nets,
Khalid for EMAN,
AMChess for BrainLearn and ShashChess, Sarona for Aurora.
The others invoked their right to be forgotten, no problem.
By uploading my
source codes on github, I was able to pass on my tools to the new generation of engine trainers.
I wish good luck to them, i remain at your disposal on Skype and in PM if needed.
Re: Train your engine !
Posted: Tue Jun 06, 2023 1:15 am
by sarona
I will register over at Infinity in order to stay in touch. That seems to be the most logical course of action. I assume it is this forum?
https://cumnor-chess-server.com/phpbb/i ... 6a8e044778
OpenChess has no PMs. Talkchess may not even be an option with all political intrigue going on there at the moment.
I know Outskirts is not a possibility either to keep in touch. Too bad, because I like what MZ has done there in dealing with trolls and leecher refugees from now defunct forums (a few of the Outskirts exiles are now at Talkchess getting on Guenther's nerves
). I like it there very much compared to the absolute zoo which were the Chess2U and ChessGod101 forums.