Proteus Chess SF - developing my Stockfish 14.1 derivative.

Mark62 wrote:

AlexChess wrote:

UPDATE 220718: ProteusSF RBE 008c with latest net , ProteusSF.exp and Eduard Nemeth's Solista 220701 bin book!

DOWNLOAD for Windows, Android and Mac M1

You cannot implement the latest network "nn-ad9b42354671.nnue" without updating the code with all the patches released

but more importantly you cannot use this ladder associated with the last green net

Code:

int scale      = 1014 + 21 * pos.non_pawn_material() / 1024;

Thank you,

I appreciate very much your suggestion!

on 009f I have changed the value to 1064 from 1092, on 008b there wasn't the same line of code due to the lack of new patches, but it is scoring very well against other derivatives on my tests. Anyway I will add patches and fix this issue asap.

In the meanwhile I'm patching the source, I'm continuing the tests, and ProteusSF RBE 008c is doing very well  

It's solid and active, using all its resources (exp and NNUE)







24 ProteusSF RBE 008c - Swordfish 15.3 games


The reality is that Stockfish code is so well written that it's resilient to doubful changes...

...Next test versus Spectral 1a that has 1064 scale value instead of 1014 of 008c

ProteusSF RBE 008c - Spectral 1.a ALL 24 PGN GAMES

DYNAMIC CHESSBOARD Spectral 1a - ProteusSF RBE 008c 0-1

DYNAMIC CHESSBOARD ProteusSF RBE 008c - Spectral 1a 1-0

DYNAMIC CHESSBOARD Spectral 1a - ProteusSF RBE 008c 1-0

FOR AVOIDING SUSPECTS ABOUT FAKE MATCHES LIKE CDR, HERE Mn/s and search depth of all engines tested until now.

And obviously I'll publish all the games with moves scores and I'm using 8moves_v3.pgn reversing sides on each position
TC: 5 MINS + 2 SECS / ALL - TABLEBASES SYZYGY 6 MAN - PONDER OFF - 64 mb CACHE - 4 CPUs on  Mac M1



Last edited by AlexChess on Wed Jul 20, 2022 3:11 pm; edited 3 times in total

New Spectral it's coming...Lot of changes my good friend.

ALL 24 PGN GAMES

Last edited by AlexChess on Wed Jul 20, 2022 3:36 pm; edited 3 times in total

ALL 24 PGN GAMES



Last edited by AlexChess on Thu Jul 21, 2022 5:26 pm; edited 13 times in total

stop playing with yourself send 008c to public.
I will test and tell you is it strong or not.

What r u on abt nw 008c is out already.

l3rilliance wrote:

stop playing with yourself send 008c to public.
I will test and tell you is it strong or not.

Don't bore MZ and me, be kind.

Oh sorry I was not trying to bore you.
I just wanted to get the download link for the 008c version and nothing else.
I have one testset with 200 positions approximately and I test engines and compare how much they solve so if you are interested in testing ProteusSF on that and have the result I can help you.
Also I dont want to further talk about MZ he knows what he did to me.

Regression tests (312 games) :

ProteusSF RBE 008c +2 13-11 /24
Blue Marlin 15.2a

ProteusSF RBE 008c +1 12.5-11.5 /24
Spectral 1a,
Brainlearn 17,
ProteusSF Aurora 009f,
Crystal 220621

ProteusSF RBE 008c = 12-12 /24
latest Stockfish 220713 ,
Stockfish Aurora 220620 ,
StockfishMZ 220622 ,
Kayra 1.5,
Sworfish 15.3 ,
Shashchess 22 ,
ProteusSF RBE 008b

ProteusSF RBE 008c -1 11.5-12,5 /24
Corchess 3 NNUE 220713

DOWNLOAD ALL 312 PGN GAMES

Last edited by AlexChess on Fri Jul 22, 2022 12:05 pm; edited 4 times in total

Ho Alex.. may I know on what time control u made ur test ?

Thank u

Zippo wrote:

Ho Alex.. may I know on what time control u made ur test ?

Thank u

5 mins + 2 secs /all 4 Mn/s on average

AlexChess wrote:

Zippo wrote:

Ho Alex.. may I know on what time control u made ur test ?

Thank u

5 mins + 2 secs /all 4 Mn/s on average

Okay thank u

buenos dias Alex! de los procesadores intel hay algunos (yo tengo un i7)que tienen una funcion llamada HYPERTHREADING en el que puedes habilitar o deshabilitar,yo he probado proteus con las dos maneras y he notado que con hiperthreading deshabilitado el modulo juega mas ràpido.tu que crees?

Mark62 wrote:

AlexChess wrote:

UPDATE 220718: ProteusSF RBE 008c with latest net , ProteusSF.exp and Eduard Nemeth's Solista 220701 bin book!

DOWNLOAD for Windows, Android and Mac M1

You cannot implement the latest network "nn-ad9b42354671.nnue" without updating the code with all the patches released

but more importantly you cannot use this ladder associated with the last green net

Code:

int scale      = 1014 + 21 * pos.non_pawn_material() / 1024;

As kindly suggested by @Mark62 , I have compiled ProteusSF RBE 008d adding latest patches and changing int scale to 1064.

Anyway 008c is very solid! -1 / 24 only against Corchess 3 NNUE 220713

+2.5 / 312 over all Stockfish derivatives

ALL DETAILS ABOUT THE REGRESSION TEST

Starting now regression tests to verify how it scores against ProteusSF RBE 008b, 008c and @Sarona 's Stockfish Aurora 220620 that, despite is still using the old net, is +6.5 points over the average of all other SF derivatives with new net!!!



If ProteusSF RBE 008d will overcome 008c on same regression test, I will release it on August, 1

Last edited by AlexChess on Fri Jul 22, 2022 6:39 pm; edited 2 times in total

dayac wrote:

buenos dias Alex! de los procesadores intel hay algunos (yo tengo un i7)que tienen una funcion llamada HYPERTHREADING en el que puedes habilitar o deshabilitar,yo he probado proteus con las dos maneras y he notado que con hiperthreading deshabilitado el modulo juega mas ràpido.tu que crees?

Strange, should be the opposite.

AlexChess wrote:

dayac wrote:

buenos dias Alex! de los procesadores intel hay algunos (yo tengo un i7)que tienen una funcion llamada HYPERTHREADING en el que puedes habilitar o deshabilitar,yo he probado proteus con las dos maneras y he notado que con hiperthreading deshabilitado el modulo juega mas ràpido.tu que crees?

Strange, should be the opposite.

hace algunos años los programadores de komodo aconsejaban deshabilitar el "hyper"

There is something wrong with last NET or tune scale settings for me and probably for @Anton101 :

ProteusSF 008d with all last patches, scale 1064 value and vdv's  nn-ad9b42354671.nnue is worse than
ProteusSF 008c without patches and 1014 scale value and same net! I stop tuning this branch and switch to Aurora testing my mods with old Sopel's nn-3c0aa92af1da.nnue until @Sarona will update it and also without NNUE at all, including Berserk 9, Seer 2.5.0, Koivisto 8.13 and Dragon 3.0 with their own NNUE instead.



ALL 008c vs 008c PGN GAMES

I'll continue all remaining 008d games against other engines (updating Spectral 2 and adding ProteusSF Aurora 009e) , but if it doesn't show much better results I'll not release it.



Last edited by AlexChess on Sun Jul 24, 2022 6:48 am; edited 1 time in total

Yes, the strength of the net nn-3c0aa92af1da.nnue of 13-05-22 is higher than the last green net nn-ad9b42354671.nnue, for this reason I returned the scale and modified the parameters of evaluation, pawns and search and found a little more force for the engine, as well as greater speed and depth.

Best regards.

l3rilliance wrote:

stop playing with yourself send 008c to public.
I will test and tell you is it strong or not.

https://rb.gy/8vo84j

I tested, this monster is stronger than the latest beta version of Stockfish.

But seem the 572MB experience did not make much different but oni eat up the memory resource of ur PC, U might as well replace it with the
experience file of Stockfish MZ if U r owning one.

In my test I will disable the EXP files because we need to test the ability of engine and no need to test how well your EXP book plays.
Chess is not formula game its a dynamic game. Not always your EXP will keep on playing its engine analysis which is more needed.
Also I have required amount of 7men tablebases to help in engine analysis. The fact also that I dont have stockfishMZ or its EXP because the engine is weak thats what testset say. There is nothing against MZ he has worked well with this engine. You also need to work with one set of code for each version you release you go to some other engine from which you copy and call yourself great programmer !? And this kid stuff you cannot stop ---> RBE ?? Rocky Balboa Edition - lol are you small kid do you see any other engine like that? you could be much better making ProteusSF 1.0 and move further with versioning. And engine version 008a, 008b, 008c which can be replaced by 8.0, 8.1, 8.2. Come on boy grow up.

Initially learn based engines appear stronger than beta of stockfish but you keep playing with it, you will see how much weak it becomes.
If learn would really help engine Stockfish authors would have implemented it. You need to test EXP and learn in longer time control like 60+15 or more. Blitz is not only time control !?

New Stockfish-dev released, based always on the same net, but build by Dubslow that has released the very good base for @Sarona 's Stockfish Aurora 220620 still stronger than all other derivatives on my tests by +6.5 /336 points on average   . I will probably update Proteus Aurora with it, leaving 008 branch, but I hope that also a new NNUE will be released in the meanwhile. Anyway 008c has a very good performance (12-12 against Spectral 2, now playing last round against Stockfish Aurora 220620 3-3 until now)

Started a new test, ProteusSF RBE 008c Classical and ProteusSF Aurora 009e Classical (NO NNUE) against the 4 best original engines with NNUE...

AlexChess wrote:

Started a new test, ProteusSF RBE 008c Classical and ProteusSF Aurora 009e Classical (NO NNUE) against the 4 best original engines with NNUE...  

Hi Alex,

You done and are doing a great job for the world chess community.

God bless you and don't stop.

Friendly, Sékou