Proteus Chess SF - developing my Stockfish 14.1 derivative.

Gator007 wrote:

Has the oo7 branch been released?

Yes

https://www.chess.com/forum/view/game-analysis/freeware-against-commercial-chess-software-arena-3-5-1-vs-chessbase-fritz-17?page=1

ProteusSF JBE 007 +3 vs Crystal 220504
Android 8 CPUS 400 Kn/s - 100 games - ProteusBook Test



ALL PGN GAMES: https://pixeldrain.com/u/fzogeDcz

AlexChess wrote:

ProteusSF JBE 007 +3 vs Crystal 220504
Android 8 CPUS 400 Kn/s - 100 games - ProteusBook Test



ALL PGN GAMES: https://pixeldrain.com/u/fzogeDcz

Many thanks Alex for your recent test. I think it is time to see an update for Crystal engine too. Also many thanks to Joseph Ellis for Crystal engine which is like a worthy diamond for chess fans. Congratulation to you for proteus victories. Whising you more and more success.
Best regards

Thanks for the link!!

DELTA TEST ProteusSF JBE 007 +45 vs Berserk 8.5.1
Android 8 CPUs 500 KN/s - ProteusBook.bin Test



ALL PGN GAMES: https://pixeldrain.com/u/MTPkC33W

Proteus like author says he has not made it from stockfish 14.1 but it is made from Shashchess. Just once create engine proteus and in first screen you can see the authors kiniama and manzo who work on shashchess. Also compare source code files with shashchess you guys can see that.
Proteus Chess SF - developing my Stockfish 14.1 derivative. - you are not doing that instead
Proteus Chess SF - developing my Shashchess 22 derivative. - this is what you are doing

You can have a look at the screenshot and get to know the truth https://we.tl/t-Vy3hPGKFyu

l3rilliance wrote:

Proteus like author says he has not made it from stockfish 14.1 but it is made from Shashchess. Just once create engine proteus and in first screen you can see the authors kiniama and manzo who work on shashchess. Also compare source code files with shashchess you guys can see that.
Proteus Chess SF - developing my Stockfish 14.1 derivative. - you are not doing that instead
Proteus Chess SF - developing my Shashchess 22 derivative. - this is what you are doing

You can have a look at the screenshot and get to know the truth https://we.tl/t-Vy3hPGKFyu

What's your point? Isn't a Shashchess derivative but a BrainLearn derivative (both are Stockfish derivatives). All them full GPLv3 compliant.

I think that very soon authors of stockfish family engines and even non stockfish engines should add this learning feature to their engines. Engines that have this option like proteus, brainlearn, shashchess, and sugar Iccf ,...play different openings better than other engines without learning feature and also they evaluate different positions much better. It seems that it will be necessary for modern engines in future.
Best regards

On the first positions of a game, even if Stockfish runs until D80 or deeper, nothing prove this game will end in 80 plies and the opponent may play different moves as those expected by the stockfish pov. So in sum, without a learning feature, the engine has no experience on the whole game, no experience on the usual endgames, no experience on the opponent playing style, no true WDL ratio, etc. Stockfish plays stronger at d80 than at d30 but if the game ends in 140 plies, his evaluation is very limited. A trained engine doesn't need d40 or deeper to outplay sf.

deeds wrote:

On the first positions of a game, even if Stockfish runs until D80 or deeper, nothing prove this game will end in 80 plies and the opponent may play different moves as those expected by the stockfish pov. So in sum, without a learning feature, the engine has no experience on the whole game, no experience on the usual endgames, no experience on the opponent playing style, no true WDL ratio, etc. Stockfish plays stronger at d80 than at d30 but if the game ends in 140 plies, his evaluation is very limited. A trained engine doesn't need d40 or deeper to outplay sf.

Following Deeds tutorial (and data) and Sarona suggestions, next Proteus version will have an even better learning feature

A trained engine :
- 500 games by opening where the engine showed weaknesses. Use few opponents that have a learning feature.
- lost games analyzed at 10min/ply when the average game's duration is about 10min, 1hr/ply when 1hr/game, etc.
- avoid q-learning at learning when you don't use q-learning at tourney else the experience data are worst than default engine. Q-learning needs thousands games by opening, it learns very very slow.

My point is there is lot of difference between stockfish 14.1 and shashchess. So tell whatever it is really based on.

l3rilliance wrote:

My point is there is lot of difference between stockfish 14.1 and shashchess. So tell whatever it is really based on.

I cannot change the title of the topic and I repeat: it's based on BrainLearn, not on shashchess .

Next build will be based on another Stockfish derivative because there are issues with BrainLearn Android compilation.

All us deserve respect and admiration for Stockfish Developers, I will always release it as GPLv3, so everyone can study source and can even

make Proteus derivatives. We can't reinvent the wheel every time. Original device authors like Grant and Mark Uniacke are stoics but they will

never compete with a very skilled community of chess engines programmers.

https://www.chess2u.com/t18076p50-proteus-chess-sf-developing-my-stockfish-14-1-derivative#114892

deeds wrote:

A trained engine :
- 500 games by opening where the engine showed weaknesses. Use few opponents that have a learning feature.
- lost games analyzed at 10min/ply when the average game's duration is about 10min, 1hr/ply when 1hr/game, etc.
- avoid q-learning at learning when you don't use q-learning at tourney else the experience data are worst than default engine. Q-learning needs thousands games by opening, it learns very very slow.

Thank you deeds. Sarona was right, your method is really good. I'm using your 250 mb Eman.exp taken from @Anton01 FB page for next Proteus version (June release) and it is doing really well!!!

Hi Alex. I use windows binary. ( bmi2) proteus plays very well and has very good evaluation (due to it’s basic brainlearn evaluation ). Now it’s very nice engine for analysis and practice. Areyou sure you want to change its brain learn base? You know better that there are many other stockfish derivatives which are very strong for engines competition but because of lacking of good and human like evaluation they are not suitable for human practice and analysis. Fortunately now proteus 007 is suitable and appropriate for both usage. I don’t want to interfere more but I think your decision needs more checking and review. May be you can solve the android version issues and it would be not necessary to change the base of engine which will cause to change its good playing style and
Evaluation.
Best regards

Homayoun wrote:

Hi Alex. I use windows binary. ( bmi2) proteus plays very well and has very good evaluation (due to it’s basic brainlearn evaluation ). Now it’s very nice engine for analysis and practice. Areyou sure you want to change its brain learn base? You know better that there are many other stockfish derivatives which are very strong for engines competition but because of lacking of good and human like evaluation they are not suitable for human practice and analysis. Fortunately now proteus 007 is suitable and appropriate for both usage. I don’t want to interfere more but I think your decision needs more checking and review. May be you can solve the android version  issues and it would be not necessary to change the base of engine which will cause to change its  good playing style and
Evaluation.
Best regards

Don't worry, the style and evaluation will be the same, the learning much better (and it will finally give a continued feedback during learning, also when saving single PV) I'm going slowly and doing a lot of tests because I want it will be really better, and with the help of Sarona, Deeds, dorsz and obviously Khalid Omar and Marco Zerbinati and Stockfish developers source it's already very competitive (just now running on Mac, Windows and Android )

AlexChess wrote:

Thank you deeds. Sarona was right, your method is really good. I'm using your 250 mb Eman.exp taken from @Anton01 FB page for next Proteus version (June release) and it is doing really well!!!

So the next Proteus will use an experience file calibrated only for the Eman's weaknesses at 3m+2s on 40t (mostly B openings). This EXP file contains Eman's evaluations that are different than BrainLearn's and have totally nothing to do with the Q-learning's. Furthermore, the experience data aren't stored/loaded in the same format/way between an experience.bin file and an eman.exp file.

Good luck to find the same thresholds/settings/logics as those used by Khalid in the Eman learning's code, even MZ didn't get the same in HypnoS. By mixing experience data like this, you replace the evaluation of Proteus by that of another hyper trained engine so you have finally admitted to us that Proteus' evaluation is way off the mark.

ANNUNCING ProteusSF RBE 008 (Rocky Balboa Edition) Release: June 2022  PLEASE DON'T ASK ME THE DOWNLOAD LINKS NOW  

PEACE WORLDWIDE!!! Only sport fights are allowed (...and mental challenges like chess are the best!   )



FREE Stockfish Derivative GPLv3 compliant based on all best SF & derivative sources (All attributions and sources will be on the zip files)

• Improved Learning feature optimized against all top engines following Deed's and Sarona's suggestions and code
  

• Learning data ProteusSF.exp expanded from shared Deed's Eman.exp 250 MB data.
  

• NNUE support
  

• Polyglot support
  

• Proteus Tournament Polyglot | PGN Books for engine-engine matches optimized against all top books
  

• Proteus Infinite Book | PGN Books (40000 ICCS games 2300-2600 ELO for long time analysis)
  

• Dorsz's Latest Unbalanced PGN Openings Test suite
  

deeds wrote:

AlexChess wrote:

Thank you deeds. Sarona was right, your method is really good. I'm using your 250 mb Eman.exp taken from @Anton01 FB page for next Proteus version (June release) and it is doing really well!!!

So the next Proteus will use an experience file calibrated only for the Eman's weaknesses at 3m+2s on 40t (mostly B openings). This EXP file contains Eman's evaluations that are different than BrainLearn's and have totally nothing to do with the Q-learning's. Furthermore, the experience data aren't stored/loaded in the same format/way between an experience.bin file and an eman.exp file.

Good luck to find the same thresholds/settings/logics as those used by Khalid in the Eman learning's code, even MZ didn't get the same in HypnoS. By mixing experience data like this, you replace the evaluation of Proteus by that of another hyper trained engine so you have finally admitted to us that Proteus' evaluation is way off the mark.

Never said the opposite. Proteus is only an EXPERIMENTAL fine tuning FULLY GPLv3 compliant of Stockfish & derivatives sources, nothing to hide.

I only know that now proteus 007 has one of the most human like evaluations like the engines sugar Iccf ,brainlearn , sashchess and crystal. These engines never overvaluate any positions resulted from different openings and also play different openings very well even without use an accessory bin book.

Homayoun wrote:

I only know that now proteus 007 has one of the most human like evaluations like the engines sugar Iccf ,brainlearn , sashchess and crystal. These engines never overvaluate any positions resulted from different openings and also play different openings very well even without use an accessory bin  book.  

So a good news for you, actual Proteus EXP 008 (dev) is based on SugaR 2.50 but the net is still the old one nn-13406b1dcbe0.nnue, since the latest isn't supported. Despite that, it is doing very well also against updated engines. I've finished 007 tests, next will be about 008 performances (only TEST games will be published, not 008 dev or final until June)

Homayoun wrote:

I only know that now proteus 007 has one of the most human like evaluations like the engines sugar Iccf ,brainlearn , sashchess and crystal. These engines never overvaluate any positions resulted from different openings and also play different openings very well even without use an accessory bin  book.  

So a good news for you, actual Proteus EXP 008 (dev) is based on SugaR 2.50 but the net is still the old one nn-13406b1dcbe0.nnue, since the latest isn't supported. Despite that, it is doing very well also against updated engines. I've finished 007 tests, next will be about 008 performances (only TEST games will be published, not 008 dev or final until June)

DELTA TEST ProteusSF JBE 007 +44 vs Fire 8NN.MC.3
Android 8 CPUs 400 KN/s - ProteusBook.bin Test[/b]



ALL PGN GAMES: https://pixeldrain.com/u/6cqrwkg8

Homayoun wrote:

I only know that now proteus 007 has one of the most human like evaluations like the engines sugar Iccf ,brainlearn , sashchess and crystal. These engines never overvaluate any positions resulted from different openings and also play different openings very well even without use an accessory bin  book.  

I have forgotten swordfish engine. It is also a fantastic engine for analysis.
(Very good evaluation and playing style)
Best regards