StockfishMZ is out

StockfishMZ 170522
Aligned with: Tune scale and optimism.
Tune scale and optimism in effort to make stockfish play more aggressively.
https://github.com/official-stockfish/Stockfish/commit/22b7909809c731aea691184dd7c1a2b02c5946af
Add SelfLearning + StockfishMZ Experience Tools + Various features added


https://pixeldrain.com/u/7NURHrKN

Great to have you here, Marco. The SugaR_xPrO and SugaR AI series of engines have been my favorites for a long time.

Look forward to testing your new creation.

sarona wrote:

Great to have you here, Marco.  The SugaR_xPrO and SugaR AI series of engines have been my favorites for a long time.

Look forward to testing your new creation.

Thanks Sarona ..
I certainly hope it is a versatile engine even if more can be done for the part that concerns the analysis
but it is a good compromise to please everyone .. (Machine room players and Iccf players)

Thanks Marco !

The arrival of an engine with his signature, is as always welcomed with applause in the community, because we know in advance, your high quality work in previous engines. At the moment I am working with StockfishMZ, I only have a doubt Mr. Zerbinati, the SugaR 2.50 experience file that I still use, is it possible to rename it and use it in this engine or will we definitely have to start from scratch? It may be a silly question, but it's an engine with your signature.

Thank you very much for your work.

Anton101 wrote:

The arrival of an engine with his signature, is as always welcomed with applause in the community, because we know in advance, your high quality work in previous engines. At the moment I am working with StockfishMZ, I only have a doubt Mr. Zerbinati, the SugaR 2.50 experience file that I still use, is it possible to rename it and use it in this engine or will we definitely have to start from scratch? It may be a silly question, but it's an engine with your signature.

Thank you very much for your work.

Hi Anton,
the logic of the learning code is compatible
however, it makes little sense to use an experience file generated from another engine's ratings
Surely more correct to create a new one and train it with the usual patience.

Yes, it makes perfect sense, that's why I thought about it a lot before I asked. But it's okay with asking. Beauty of work! Regular training, perfect! Thank you so much for coming back with this great gift. Best wishes!

From Peter Martan (CSS-Forum)
https://forum.computerschach.de/cgi-bin/mwf/topic_show.pl?pid=155057

Hi friend, impossible to download StockfishMZ

Can you create another link?

Wait a moment, MZ is fixing bugs...

deeds wrote:

Wait a moment, MZ is fixing bugs...

Thanks very much my friend.

Mark62 wrote:

From Peter Martan (CSS-Forum)
https://forum.computerschach.de/cgi-bin/mwf/topic_show.pl?pid=155057

Code:

  Program                                    Elo   +/-  Matches  Score   Av.Op.   S.Pos.   MST1    MST2   RIndex

  1 StockfishMZ170522avx2-ME-12-PE+12        : 3592    7   2522    65.2 %   3483    95/128    3.1s    6.2s   0.73
  2 Crystal251221                            : 3589    7   2544    64.6 %   3485    93/128    2.8s    6.1s   0.62
  3 Swordfish15.1-avx2                       : 3586    7   2476    64.5 %   3483    92/128    2.9s    6.3s   0.69
  4 ShashChess22-GoldDigger+Tal              : 3586    7   2537    64.4 %   3483    95/128    3.3s    6.3s   0.64
  5 CorChess3230422                          : 3583    7   2451    64.0 %   3483    92/128    2.9s    6.3s   0.63
  6 Stockfish15                              : 3573    7   2354    62.4 %   3485    87/128    3.0s    6.8s   0.63
  7 Stockfish140522                          : 3572    7   2364    62.2 %   3486    87/128    3.4s    7.1s   0.61
  8 Ceres0.97RC3-ap-mish-2000000             : 3568    7   2349    61.5 %   3486    83/128    3.1s    7.3s   0.62
  9 Ceres0.97RC3-783257                      : 3564    8   2234    60.5 %   3489    77/128    2.5s    7.5s   0.62
10 StockfishMZ170522avx2                     : 3562    7   2340    60.9 %   3485    84/128    3.3s    7.3s   0.58
11 Lc0v0.30.0-dev-mish-2000000               : 3553    8   2180    58.9 %   3490    74/128    2.7s    7.9s   0.59
12 Lc0v0.30.0-dev-783272                     : 3550    8   2160    58.4 %   3491    71/128    2.4s    8.0s   0.60
13 Dragon3byKomodoChess64-bit                : 3538    8   2181    56.3 %   3494    72/128    3.8s    8.7s   0.50
14 Koivisto8.8                               : 3447    9   1712    40.4 %   3514    37/128    3.6s   11.7s   0.30
15 Rebel15x2                                 : 3443    9   1707    39.6 %   3516    35/128    3.1s   11.7s   0.27
16 RubiChess20220223(bmi2)                   : 3432    9   1758    38.2 %   3515    38/128    5.8s   12.3s   0.20
17 Houdini4Prox64                            : 3422   10   1691    36.0 %   3522    30/128    4.0s   12.4s   0.18
18 Stockfish1164POPCNT                       : 3414    9   1692    35.1 %   3520    32/128    5.6s   12.7s   0.14
19 Seer2.5.0                                 : 3403   10   1612    33.3 %   3524    27/128    4.6s   12.8s   0.18
20 ZappaMexicoIIx64                          : 3391   10   1556    30.9 %   3530    20/128    2.2s   13.0s   0.20
21 DeepHIARCS15.0                            : 3374   10   1537    29.0 %   3529    19/128    4.1s   13.4s   0.16
22 rofChade3.0BMI2AVX2                       : 3371    9   1555    29.0 %   3527    21/128    5.3s   13.4s   0.13
23 Fritz15                                   : 3367   10   1534    28.1 %   3531    18/128    4.8s   13.6s   0.15
24 Fritz17Popcnt                             : 3356   10   1516    26.6 %   3532    15/128    3.8s   13.7s   0.13
25 DeepShredder13x64                         : 3348   10   1522    25.8 %   3532    17/128    6.3s   13.8s   0.08
26 Wasp5.50                                  : 3346   10   1492    25.4 %   3533    14/128    4.3s   13.8s   0.13

MST1  : Mean solution time (solved positions only)
MST2  : Mean solution time (solved and unsolved positions)
RIndex: Score according to solution time ranking for each position


The upper half of the list is much better as for rankings then the lower one as error bars show too, more than this too little numbers of solved positions with this hardware- time make the ranking less reliable per se.
I just give the whole list to show the numbers of listed engines because it's a measurement for the number of matches which are taken into account as for rankings and ratings.

EloStatTS doesn't count solved positions primarily, it compares each one time of solution of each position solved and unsolved of each one single engine to each one of all of the others, just like it does with results of engine- engine- games with EloStat from same author.

StockfishMZ- setting marked with ME-12-PE+12 has Materialistic Evalution Strategy set to minium, Positional Evaluation Strategy to maximum (+12), which makes a big difference here: -12+12-setting leading, default setting ranked as nr.10.
So it seems this one setting is tuned much better as for this one special test, at least as for this one NNUE (tried ShashChess and BlueMarlin already one more time too in meantime with nn-3c0aa92af1da.nnue, getting significantly less numbers of solved positions with both of them compared to the one but last green SF dev- net nn-d0b74ce1e5eb.nnue, they are still listed with above).

Thanks a lot for the new engine- branch replacing SugaR AI:
Peter

Well done finishing ahead of Crystal (and finishing first overall with modified settings) with this particular testsuite.  That is an excellent result posted from a highly respected tester, Peter Martan.

He occasionally posts at over at Talkchess, too. It was he that drew my interest in using EloStatTS (I use 1.04).

Just an experiment !

The new StockfishMZ works perfectly with the huge Eman.exp (renamed StockfishMZ.exp). Promising.

Started !

Both engines play with Eman.exp. I'm curious how they will modify-each engine-Eman.exp (StockfishMZ.exp).

RSilvian wrote:

Started !

Both engines play with Eman.exp. I'm curious how they will modify-each engine-Eman.exp (StockfishMZ.exp).

Same net ?

deeds wrote:

RSilvian wrote:

Started !

Both engines play with Eman.exp. I'm curious how they will modify-each engine-Eman.exp (StockfishMZ.exp).

Same net ?

NO ! The default ones. SF-nn-3c0aa92af1da.nnue for StockfishMZ & SF-nn-6877cd24400e.nnue for EMAN 7.93. The same are the experience files. Interesting match.

The installation folder (SFMZ) looks like:

StockfishMZ170522avx2-ME-12-PE+12 what means This Me -12 - PE+12 please can some one explain me

coci wrote:

StockfishMZ170522avx2-ME-12-PE+12  what means This Me -12 - PE+12 please can some one explain me

thank you

coci wrote:

StockfishMZ170522avx2-ME-12-PE+12  what means This Me -12 - PE+12 please can some one explain me

Materialistic Evaluation Strategy, default 0, min -12, max 12
Positional Evaluation Strategy, default 0, min -12, max 12

ME-12 = Materialistic Evaluation Strategy -12 = min materialistic
PE+12 = Positional Evaluation Strategy 12 = max positional

So the engine doesn't care about the material on the board and searches for the best positional play = perfect settings for solving strategic test suites.
Of course with these settings, the scores of the StockfishMZ engine have nothing to do with the scores of the sfnnv5 net, and again less to do with the scores of Eman from the experience file.
It's very weird to use these "test suites" settings in a tourney where the goal is to predict the most opponent's moves in order to win the game, not to find the expected move of a test suite.

coci wrote:

StockfishMZ170522avx2-ME-12-PE+12  what means This Me -12 - PE+12 please can some one explain me

1) Materialistic Evaluation Strategy: Minimum = -12, Maximum = +12, Default = 0. Lower values will cause the engine assign less value to material differences between the sides. More values will cause the engine to assign more value to the material difference.
2) Positional Evaluation Strategy: Minimum = -12, Maximum = +12, Default = 0. Lower values will cause the engine assign less value to positional differences between the sides. More values will cause the engine to assign more value to the positional difference.

The default value for both options (0 = zero) is equivalent to the default evaluation strategy of Stockfish.
khalid-a-omar authored

I am modifying some bugs and separating the functions
I will create an engine for playing games and one for hard positions analysis and solutions

Well, my match test SFMZ vs. Eman 7.93 will have a first half with the settings: ME-12/PE+12 and a second half with the default settings (ME-0/PE-0). I'm curious to see the influence on the result.