Andscacs nnue 0.1

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

Moderator: Ras

User avatar
cdani
Posts: 2204
Joined: Sat Jan 18, 2014 10:24 am
Location: Andorra

Andscacs nnue 0.1

Post by cdani »

Hello!
I was curious about what could I obtain adding nnue to Andscacs.
I went straight to nnue probe of Daniel Shawul:
https://github.com/dshawul/nnue-probe
After some work, manual tuning and testing, the result was beyond my expectations:

Code: Select all

50+0.5 
   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Andscacs 0.95183    : 3042.4   14.5    620.0     695   89.2%
   2 Andscacs 0.95143    : 2669.6   14.5     75.0     695   10.8%
So more than 350 elo in self play!

Here the Windows executable and latest Stockfish nnue net:
https://1drv.ms/u/s!Aj8wlSmr_Hq3islynT1 ... Q?e=FVH5CN
You can change the nnue net, just overwrite the nn.bin file.

Of course rating lists that care about engine originality should not add this version.

Is a bit far from Stockfish, but is not bad:

Code: Select all

500+2
   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Stockfish 311220    : 2961.6   20.8     83.5     109   76.6%
   2 Andscacs 0.95183    : 2750.4   20.8     25.5     109   23.4%
AndrewGrant
Posts: 1953
Joined: Tue Apr 19, 2016 6:08 am
Location: U.S.A
Full name: Andrew Grant

Re: Andscacs nnue 0.1

Post by AndrewGrant »

cdani wrote: Sun Jan 17, 2021 7:55 pm Hello!
I was curious about what could I obtain adding nnue to Andscacs.
I went straight to nnue probe of Daniel Shawul:
https://github.com/dshawul/nnue-probe
After some work, manual tuning and testing, the result was beyond my expectations:

Code: Select all

50+0.5 
   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Andscacs 0.95183    : 3042.4   14.5    620.0     695   89.2%
   2 Andscacs 0.95143    : 2669.6   14.5     75.0     695   10.8%
So more than 350 elo in self play!

Here the Windows executable and latest Stockfish nnue net:
https://1drv.ms/u/s!Aj8wlSmr_Hq3islynT1 ... Q?e=FVH5CN
You can change the nnue net, just overwrite the nn.bin file.

Of course rating lists that care about engine originality should not add this version.

Is a bit far from Stockfish, but is not bad:

Code: Select all

500+2
   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Stockfish 311220    : 2961.6   20.8     83.5     109   76.6%
   2 Andscacs 0.95183    : 2750.4   20.8     25.5     109   23.4%
:( I'll stick to plain old Andscacs. There are more than enough Stockfish nets floating around.
User avatar
cdani
Posts: 2204
Joined: Sat Jan 18, 2014 10:24 am
Location: Andorra

Re: Andscacs nnue 0.1

Post by cdani »

AndrewGrant wrote: Sun Jan 17, 2021 8:03 pm :( I'll stick to plain old Andscacs. There are more than enough Stockfish nets floating around.
Was just a fun experiment :-) Nothing bad.
Damir
Posts: 2865
Joined: Mon Feb 11, 2008 3:53 pm
Location: Denmark
Full name: Damir Desevac

Re: Andscacs nnue 0.1

Post by Damir »

Thanks for the new Andscacs Dani. Good to see you start working on your engine again. :) :)
User avatar
maksimKorzh
Posts: 775
Joined: Sat Sep 08, 2018 5:37 pm
Location: Ukraine
Full name: Maksim Korzh

Re: Andscacs nnue 0.1

Post by maksimKorzh »

cdani wrote: Sun Jan 17, 2021 7:55 pm Hello!
I was curious about what could I obtain adding nnue to Andscacs.
I went straight to nnue probe of Daniel Shawul:
https://github.com/dshawul/nnue-probe
After some work, manual tuning and testing, the result was beyond my expectations:

Code: Select all

50+0.5 
   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Andscacs 0.95183    : 3042.4   14.5    620.0     695   89.2%
   2 Andscacs 0.95143    : 2669.6   14.5     75.0     695   10.8%
So more than 350 elo in self play!

Here the Windows executable and latest Stockfish nnue net:
https://1drv.ms/u/s!Aj8wlSmr_Hq3islynT1 ... Q?e=FVH5CN
You can change the nnue net, just overwrite the nn.bin file.

Of course rating lists that care about engine originality should not add this version.

Is a bit far from Stockfish, but is not bad:

Code: Select all

500+2
   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Stockfish 311220    : 2961.6   20.8     83.5     109   76.6%
   2 Andscacs 0.95183    : 2750.4   20.8     25.5     109   23.4%
Hmm... 350 Elo doesn't seem to much - I had around 600 Elo gain in BBC which is very very basic but stil 2900+ at CCRL:
https://www.computerchess.org.uk/ccrl/4 ... NUE_64-bit

I think if you tune eval pruning and make 100% sure to give 50rule move penalty in eval() it would be much stronger.
Without 50rule penalty SF NNUE failed to find checkmate in KQPK endgame in my case.

Anyway congrats with your experiment - this is obviously fun!
I wish someone created a very detailed tutorial on how to design/implement/train NNUE from scratch -
unfortunately there's a abyss gap on this question - those smart enough just doing their own networks
while noobs are doomed to HCE. For me this is discrimination.
I would've definitely create a youtube tutorial on NNUE from scratch if only I was smart enough(((
User avatar
Sylwy
Posts: 4807
Joined: Fri Apr 21, 2006 4:19 pm
Location: IAȘI - the historical capital of MOLDOVA
Full name: Silvian Rucsandescu

Re: Andscacs nnue 0.1

Post by Sylwy »

THANK YOU ! :wink:

Image

Frank Quisinsky
Posts: 6888
Joined: Wed Nov 18, 2009 7:16 pm
Location: Gutweiler, Germany
Full name: Frank Quisinsky

Re: Andscacs nnue 0.1

Post by Frank Quisinsky »

Hi Daniel,

a good start point with a new motivation!
Better is to develops an own net and maybe your site will be again available.

I am very happy that you are back.

If you create a new website, I have a Andscacs collection for you :-)
You can add a link to my site or you can add your own versions (the collection by Andscacs lovers) on your own site.
In this case I will delete the *.zp file from my site ... of course, your decision.

Go Andscacs go ...
Great engine, great playing style and the programmer is back ... great news!

Best for you
Frank
User avatar
Gabor Szots
Posts: 1433
Joined: Sat Jul 21, 2018 7:43 am
Location: Budapest, Hungary
Full name: Gabor Szots

Re: Andscacs nnue 0.1

Post by Gabor Szots »

maksimKorzh wrote: Sun Jan 17, 2021 9:36 pm Hmm... 350 Elo doesn't seem to much - I had around 600 Elo gain in BBC which is very very basic but stil 2900+ at CCRL:
https://www.computerchess.org.uk/ccrl/4 ... NUE_64-bit
You went from 2100 to 2900. He went from 3200 to 3450 (approximately).
Gabor Szots
CCRL testing group
User avatar
maksimKorzh
Posts: 775
Joined: Sat Sep 08, 2018 5:37 pm
Location: Ukraine
Full name: Maksim Korzh

Re: Andscacs nnue 0.1

Post by maksimKorzh »

Gabor Szots wrote: Sun Jan 17, 2021 10:45 pm
maksimKorzh wrote: Sun Jan 17, 2021 9:36 pm Hmm... 350 Elo doesn't seem to much - I had around 600 Elo gain in BBC which is very very basic but stil 2900+ at CCRL:
https://www.computerchess.org.uk/ccrl/4 ... NUE_64-bit
You went from 2100 to 2900. He went from 3200 to 3450 (approximately).
Oh... Should've checked the rating firsts...
In this case very very impressive.
Dann Corbit
Posts: 12777
Joined: Wed Mar 08, 2006 8:57 pm
Location: Redmond, WA USA

Re: Andscacs nnue 0.1

Post by Dann Corbit »

cdani wrote: Sun Jan 17, 2021 7:55 pm Hello!
I was curious about what could I obtain adding nnue to Andscacs.
I went straight to nnue probe of Daniel Shawul:
https://github.com/dshawul/nnue-probe
After some work, manual tuning and testing, the result was beyond my expectations:

Code: Select all

50+0.5 
   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Andscacs 0.95183    : 3042.4   14.5    620.0     695   89.2%
   2 Andscacs 0.95143    : 2669.6   14.5     75.0     695   10.8%
So more than 350 elo in self play!

Here the Windows executable and latest Stockfish nnue net:
https://1drv.ms/u/s!Aj8wlSmr_Hq3islynT1 ... Q?e=FVH5CN
You can change the nnue net, just overwrite the nn.bin file.

Of course rating lists that care about engine originality should not add this version.
This is not only a dreaded disease, it is stupidity at its very highest leve.
Please explain to me how letting a bunch of engines mindlessly pound away at each other and writing out a bunch of numbers that nobody even knows exactly what they are expresses originality.

In the first place, it is well demonstrated in copyright law that a table of numbers (e.g. an NN net) does not represent originality and cannot be copyrighted.

In the second place, the humans are adding no additional originality whatsoever. It is like there are a million little workers all smashing a block of marble into dust and at the end, someone says that they invented dust.

I see very little originality in any of this, and yet it is being used as an excuse to exclude bands of people and engines from tournaments.
I guess they did not copy as quickly as the others who are accepted.

The chess community can have their phony-baloney "This is my original stuff and you had better not touch it" mentality. i find it very tiring,

Do not get me wrong. It is a lot of work to write LC0. It is a lot of work to graft NNUE onto SF. It is a lot of work to run a million trials to make a net.
It is just that there is no originality in it whatsoever.
Is a bit far from Stockfish, but is not bad:

Code: Select all

500+2
   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Stockfish 311220    : 2961.6   20.8     83.5     109   76.6%
   2 Andscacs 0.95183    : 2750.4   20.8     25.5     109   23.4%
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.