Senpai 3.0 by Fabien LETOUZEY is available!

[Ryan Benitez]

Hi there,

Senpai 3.0 by Fabien LETOUZEY (France) is released on my older website.

https://www.amateurschach.de (download selection)
https://www.amateurschach.de/main/_download.htm
https://www.amateurschach.de/download/_ ... rsions.zip (12.072Mb ... direct download link).

Inside are
- Senpai 1.0, Senpai 2.0, the "older" files from my site.
- Senpai 3.0 JA compiles, sources, readme, copyright.

Jim compiled Senpai 3.0 RC1 in secret mission, I asked him.
Thank you for your time again Jim!



RC1 is now the release version!!

If Jim read that.
Jim, you can also public your Senpai compiles with the sources for your repo.
Please switch only the eval and readme file.

On an Intel i9 10900k I test with 8+2.
I believe, not sure (I have 2.000 games vs. 20 opponents only) ... Elo is a bit more as +200 to Senpai 2.0
Around the same playing strength Shredder 13 have.

Senpai 2.0 participant in FCP-Tourney-2020 ...
Can be see better ...

Code: Select all

   # Player                           :      Elo  Games  Score%   won  draw  lost  Points  Draw%   Error   OppAvg   OppE  MoveAvg
  -------------------------------------------------------------------------------------------------------------------------------
  01. Stockfish 11 BMI2 x64           :  3341.82  2.000    83.7  1353   643     4  1674.5   32.1   13.71  3029.12  10.38     81.1
  02. Komodo 14.0 BMI2 x64            :  3271.29  2.000    77.7  1150   809    41  1554,5   40.5   12.57  3030.88  10.41     78.7
  03. Houdini 6.03 Pro x64            :  3261.82  2.000    76.8  1104   865    31  1536,5   43.3   11.55  3031.12  10.44     81.0
  04. Ethereal 12.25 PEXT x64         :  3202.24  2.000    70.7   909  1009    82  1413,5   50.5   10.71  3032.61  10.46     93.0
  05. SlowChess BC 2.2 x64            :  3171.54  2.000    67.2   843  1002   155  1344,0   50.1   10.06  3033.37  10.48     84.0
  06. Fire 7.1 POP x64                :  3169.39  2.000    67.0   818  1042   140  1339,0   52.1   11.17  3033.43  10.45     86.2
  07. Xiphos 0.6 BMI2 x64             :  3147.47  2.000    64.3   755  1064   181  1287,0   53.2   10.57  3033.98  10.46     84.0
  08. Booot 6.4 POP x64               :  3147.06  2.000    64.3   733  1106   161  1286,0   55.3   10.41  3033.99  10.47     77.5
  09. rofChade 2.3 BMI x64            :  3144.36  2.000    64.0   764  1031   205  1279,5   51.5   10.37  3034.05  10.47     89.6
  10. Laser 1.7 BMI2 x64              :  3120.35  2.000    61.0   672  1097   231  1220,5   54.9   10.05  3034.65  10.48     85.8
  -------------------------------------------------------------------------------------------------------------------------------
  11. Shredder 13 x64                 :  3105.01  2.000    59.1   656  1052   292  1182,0   52.6   10.69  3035.04  10.46     85.8
  12. Fritz 17 (Ginkgo) x64           :  3104.82  2.000    59.1   635  1093   272  1181,5   54.6    9.75  3035.04  10.48     85.6
  13. Schooner 2.2 SSE x64            :  3102.45  2.000    58.8   606  1139   255  1175,5   57.0   10.41  3035.10  10.47     86.0
  14. Defenchess 2.2 POP x64          :  3095.96  2.000    58.0   586  1146   268  1159,0   57.3    9.91  3035.26  10.48     91.5
  15. RubiChess 1.7.3 x64             :  3092.23  2.000    57.5   611  1077   312  1149,5   53.9   10.19  3035.36  10.47     88.2
  16. Andscacs 0.95 BMI2 x64          :  3084.43  2.000    56.5   589  1081   330  1129,5   54.0   10.10  3035.55  10.48     84.4
  17. Fizbo 2.0 BMI2 x64              :  3068.55  2.000    54.4   582  1013   405  1088,5   50.6    9.98  3035.95  10.48     88.5
  18. Arasan 22.0 BMI2 x64            :  3038.06  2.000    50.5   490  1038   472  1009,0   51.9   10.25  3036.71  10.47     86.4
  19. GullChess 3.0 BMI2 x64          :  3035.00  2.000    50.0   451  1100   449  1001,0   55.0    9.87  3036.79  10.48     89.6
  20. Fritz 16 (Rybka) x64            :  2998.27  2.000    45.3   397  1016   587   905,0   50.8    9.78  3037.71  10.48     87.7
  -------------------------------------------------------------------------------------------------------------------------------
  21. Pedone 2.0 BMI2 x64             :  2990.00  2.000    44.2   362  1043   595   883,5   52.1    9.91  3037.91  10.48     88.1
  22. Chiron 4 x64                    :  2987.88  2.000    43.9   358  1040   602   878,0   52.0    9.93  3037.97  10.48     88.9
  23. Vajolet2 2.8 BMI2 x64           :  2986.15  2.000    43.7   345  1057   598   873,5   52.9    9.79  3038.01  10.48     88.5
  24. Winter 0.8 x64                  :  2985.57  2.000    43.6   391   962   647   872,0   48.1   10.34  3038.02  10.47     85.0
  25. Wasp 4.00 Modern x64            :  2984.99  2.000    43.5   342  1057   601   870,5   52.9    9.69  3038.04  10.49     88.6
  26. Critter 1.6a x64                :  2974.14  2.000    42.1   353   979   668   842,5   49.0   10.35  3038.31  10.47     89.8
  27. Igel 2.5.0 BMI2 x64             :  2971.80  2.000    41.8   314  1045   641   836,5   52.3    9.99  3038.37  10.48     83.1
  28. Equinox 3.30 x64                :  2971.61  2.000    41.8   305  1062   633   836,0   53.1   10.03  3038.37  10.48     90.4
  29. Nirvanachess 2.4 POP x64        :  2967.71  2.000    41.3   300  1052   648   826,0   52.6   10.03  3038.47  10.48     85.8
  30. Nemorino 5.00 BMI2 x64          :  2954.56  2.000    39.6   323   939   738   792,5   47.0   10.44  3038.80  10.47     89.9
  -------------------------------------------------------------------------------------------------------------------------------
  31. iCE 4.0 v853 Modern x64         :  2945.84  2.000    38.5   292   957   751   770,5   47.9   10.05  3039.02  10.48     85.4
  32. Demolito 2020-05-14 PEXT x64    :  2945.44  2.000    38.5   311   917   772   769,5   45.9   10.53  3039.03  10.46     96.4
  33. Protector 1.9.0 x64             :  2938.85  2.000    37.6   264   978   758   753,0   48.9   10.41  3039.19  10.47     88.0
  34. Hannibal 1.7 x64                :  2934.63  2.000    37.1   260   965   775   742,5   48.3   10.46  3039.30  10.47     91.1
  35. Texel 1.07 BMI2 x64             :  2931.60  2.000    36.8   263   944   793   735,0   47.2   10.27  3039.37  10.47     89.5
  36. Minic 2.33 x64                  :  2929.38  2.000    36.5   269   921   810   729,5   46.0   10.56  3039.43  10.46     84.5
  37. Senpai 2.0 BMI2 x64             :  2910.10  2.000    34.1   220   925   855   682,5   46.3   10.62  3039.91  10.46     88.9
  38. Combusken 1.2.0 x64             :  2896.88  2.000    32.5   204   894   902   651,0   44.7   10.25  3040.24  10.47     86.1
  39. SmarThink 1.98 AVX2 x64         :  2886.62  2.000    31.4   230   794   976   627,0   39.7   10.80  3040.50  10.46     86.9
  40. Monolith 2 PEXT x64             :  2858.99  2.000    28.2   122   885   993   564,5   44.3   11.01  3041.19  10.45     83.0
  -------------------------------------------------------------------------------------------------------------------------------
  41. Rodent IV 0.22 POP x64          :  2851.68  2.000    27.4   129   839  1032   548,5   42.0   11.53  3041.37  10.44     86.9
  -------------------------------------------------------------------------------------------------------------------------------

Best
Frank

Thank you Frank and of course thank you Fabien!

[Frank Quisinsky]

Shortly, no lose on time games ...
Shredder 12 GUI, Windows 11 Pro, i9 10900k, 256Mb Hash with 8+2 after the 2.000 test games.
No problems here!

No reason to say thanks to me.
I am the slave of the engines, not more.
Not many work for me to added it on my older site.

But Jim did a great job.
He found a bit during his work and provided great assistance!!
Hm, sorry ... after all I understand.

Best
Frank

[Frank Quisinsky]

Hi there,

in my humble opinion ...

HCE = Hand-Crafted-Evaluation
NN = Neural-Network

That can't be all today!??

The correct name of Senpai 3.0 can be ...
Senpai 3.0 ML JA

Means ..
ML = Machine-Learning

If so, the question is what other programmer do here, means how many others based on ML.
Thats not my topic and I have not enough knowledge about it to speak about it.

The problem is ...
My own engine-overview.
To speak about NN or HCE only is perhaps absolutely wrong.
Programmer saw that and laugh about it, for that reason that I can't evaulate it with the knowledge I have about it.

But after the mails with Fabien I think today ... HCE / NN can't be all.

The term HCE is too broad. Is everything that isn't NN considered HCE?
I'm overwhelmed thinking about the details.

HCE can't be ML ... strictly speaking, or?

Best
Frank

[Xann]

A CEGT tester friend reports that Senpai loses on time in at least 50 percent of its games.

Yesterday I noticed that the internal 'move overhead' was only 0.01s instead of 0.1s that I thought I used. I think that's the culprit.

It needs a combination of factors to cause trouble I'm guessing, such as:
- repeating time control
- slow GUI
- Senpai in a bad position on the board (strong opponents?)

It cannot be fixed without a recompilation, and therefore a change in version number?

Fabien.

[Xann]

Yesterday I noticed that the internal 'move overhead' was only 0.01s instead of 0.1s that I thought I used. I think that's the culprit.

It needs a combination of factors to cause trouble I'm guessing, such as:
- repeating time control
- slow GUI
- Senpai in a bad position on the board (strong opponents?)

It cannot be fixed without a recompilation, and therefore a change in version number?

Fabien.

If my theory is correct, the two '1E-2' in src/engine/search.cpp should be replaced with '0.1'.

Unfortunately that's a fix that only programmers can try.

Fabien.

[Xann]

Another long-time and respected tester asks:

I wonder if the commas in the uci options might confuse some guis (and how did they appear anyway? Fabien never had them before in Fruit/Senpai)?

Code: Select all

id name Senpai 3.0
id author Fabien Letouzey
option name Hash type spin default 64, min 16, max 32768
option name Threads type spin default 1, min 1, max 64
option name Draw Score type spin default 0, min -100, max 100
option name UCI_Chess960 type check default false
uciok

Actually, there is not one single uci engine so far which does this.

Well spotted, that's another mistake. The commas in the source code make sense to Rust outside of the quotes, but not in the UCI text.

Fabien.

[Graham Banks]

Great to see you back regardless, Fabien.

I've always wondered if you're the same Xann that's well-known in Football Manager circles?

[Xann]

Great to see you back regardless, Fabien.

I've always wondered if you're the same Xann that's well-known in Football Manager circles?

No, I am not.

I made up the name when I was an Amiga demo programmer, as was customary. Unfortunately I learned much later that it is a female first name, common in Asia.

I contacted a 'xann.com' company once about the name, and I learned that it was the name of the secretary. I like wondering about how that happened.

Fabien.

[Xann]

Yesterday I noticed that the internal 'move overhead' was only 0.01s instead of 0.1s that I thought I used. I think that's the culprit.

It needs a combination of factors to cause trouble I'm guessing, such as:
- repeating time control
- slow GUI
- Senpai in a bad position on the board (strong opponents?)

It cannot be fixed without a recompilation, and therefore a change in version number?

Fabien.

If my theory is correct, the two '1E-2' in src/engine/search.cpp should be replaced with '0.1'.

Unfortunately that's a fix that only programmers can try.

Fabien.

To clarify, this is the situation.

If we fix the 'move overhead', with the value that I intended for the release, the engine will allocate slightly less time for each move. That means that even tests that don't have any problem with time will no longer correspond to the fixed version (which I've named 3.0+ on my computer). It would be nice if removing the UCI commas solved everything, but ...

I assume that this kind of situation has happened before.

It is the middle of the night in Europe, where Frank and Jim live (me too, but I have sleeping problems). I think it's fair to say that it won't be fixed soon. Let's wait for an opportunity to check that the change actually solves the problem.

I am saying this so that testers don't lose time running a version that won't make the rating lists in the end.

Fabien.

[Xann]

I can produce a Windows executable on my Mac. Not JA caliber, but it will be enough to investigate the time losses.

If someone who experiences the problem could contact me, that would be appreciated. I think it has to be by email, because I need to send a file.

Fabien (xann@tutamail.com).