Honey XR7 and company near release..

[MikeB]

The build script mow includes 10 engines with different styles off play and analysis. On my Mac Pro, the 10 engines are built in 5 minutes,

They all have been updated to the latest current dev-Stockfish source

Code: Select all

### Bench signatures (total nodes searched using bench)
### Based on commits through 11/25/2019:
### ======================================================
### Black-Diamond-XR7.nodes:Nodes searched  : 9311931
### Blue-Honey-FD-XR7.nodes:Nodes searched  : 13276817
### Blue-Honey-XR7.nodes:Nodes searched  : 5455321
### Bluefish-FD-XR7.nodes:Nodes searched  : 12051099
### Bluefish-XR7.nodes:Nodes searched  : 5589824
### Honey-FD-XR7.nodes:Nodes searched  : 11434545
### Honey-XR7.nodes:Nodes searched  : 5092260
### Stockfish-FD-XR7.nodes:Nodes searched  : 10845732
### Stockfish-XR7.nodes:Nodes searched  : 5742013
### Weakfish-XR7.nodes:Nodes searched  : 8134785
### ====================================================== 

New this release is 'Black Diamond' - which is essentially Crystal by Joe Ellis with the bells and whistles for human play including , Play by FIDE Elo , 4 opening books, improved adaptive play , improved variety play . At full strength , is about 50 to 60 Elo weaker than Stockfish - which is not bad considering all it can do , thanks to Joe Ellis.

Also new in this release is "Weakfish" , which is essentially Stockfish with all the search Elo gainers removed, but the evaluation code left intact. Plays about 2400 FIDE full strength and maybe around 730 FIDE at the weakest setting of FIDE 1000.

Bluefish, another SF derivative, was authored originally by Eelco de Groot and I have agreed to keep it current more or less with dev-Stockfish.
Blue-Honey is a blend of Honey and Bluefish. Generally about 30-40 Elo weaker than SF

Honey is the continuation of McBrain, etc. usually slightly weaker than cur-dev-SF.

FR denotes that those versions also have Joe Ellis' Fortress Detection code from Crystal added in. FD versions are usually about 30 to 40 Elo weaker than the standard versions.

One single source, one makefile, one build script, builds all 10 engines - all updated to the latest current-dev-Stockfish all with the bells and whistles of Honey - centipawn output is now the default , but Scoring % is still an option. They all have Play by Elo, adaptive play, variety play and keyboard shortcuts when using an engine at the terminal. Also a new play by NPS feature was added , levels 1 to 60, starts at about 50 nps and goes to about 150K nps., roughly doubling NPS every 5 levels. Only Honey is calibrated to CCRL and FIDE, recalibrated with this release, the other engines just use the same code that was calibrated for Honey. At best , calibration is a little rough - very hard to do.

Running a final test now just to make sure nothing is broken - nothing is broken is defined as Honey being within 10 Elo or so of curr-dev-SF.

Follow along here if you wish:

Results:
https://www.dropbox.com/s/txmawp4l26te4 ... a.txt?dl=0

PGN:
https://www.dropbox.com/s/9oi23puprwe85zf/all.txt?dl=0

500 games at tc 5 min with 3 sec increment (about 13 hours or so):

Early results:

Code: Select all

20 game(s) loaded
Rank Name                      Rating   Δ     +    -     #     Σ    Σ%     W    L    D   W%    =%   OppR 
---------------------------------------------------------------------------------------------------------
   1 Honey-XR7                  1069   0.0   54   54    20   11.0  55.0    3    1   16  15.0  80.0  1039 
   2 Stockfish 251119 64 POPC   1039  29.4   54   54    20    9.0  45.0    1    3   16   5.0  80.0  1069 
---------------------------------------------------------------------------------------------------------
20 game(s) loaded

Hopefully a release later this week.

also - "XR7":

'X' based on Stockfish 10+
'R' release
'7' release #7 under 'X'

[Marek Soszynski]

Michael, you have already done so much for us, so on the principle of If you want a job done ask a busy person, as opposed to If you want a difficult job done give it to a lazy person, I wonder whether you couldn't make a Vanilla Honey. Mm...

I use engines almost exclusively for analysis; not for tournaments or game-playing. Many engine parameters are simply irrelevant for my purposes. I don't need Book options or Ponder or Skill Level, for example. And I wonder whether excluding these from the code might speed the engine up, however little. And now here's the vanilla bomb...

I very often use engines in parallel, set to a single thread each. So I wonder whether you couldn't make a Raw Honey, again without book options, etc., but without SMP, geared for single thread use, again in the hope of a speed up, however little.

[MikeB]

Michael, you have already done so much for us, so on the principle of If you want a job done ask a busy person, as opposed to If you want a difficult job done give it to a lazy person, I wonder whether you couldn't make a Vanilla Honey. Mm...

I use engines almost exclusively for analysis; not for tournaments or game-playing. Many engine parameters are simply irrelevant for my purposes. I don't need Book options or Ponder or Skill Level, for example. And I wonder whether excluding these from the code might speed the engine up, however little. And now here's the vanilla bomb...

I very often use engines in parallel, set to a single thread each. So I wonder whether you couldn't make a Raw Honey, again without book options, etc., but without SMP, geared for single thread use, again in the hope of a speed up, however little.

Shouldn’t be that hard actually. I was actually doing that at one point , but I didn’t maintain it. If I do fix it , will you promise that you will learn how to compile your own ? Rather make this a self service operation.

[Marek Soszynski]

If I do fix it , will you promise that you will learn how to compile your own ? Rather make this a self service operation.

Thanks Mike. I promise.

[MikeB]

If I do fix it , will you promise that you will learn how to compile your own ? Rather make this a self service operation.

Thanks Mike. I promise.

I just fixed the code so that all the add'l features can be turned off at compile time if one desires. As part of that process, since these are compile time options there was no measurable difference from inactivating the features you found undesirable. But the upside is that I fixed the fast magics code that is part of the additional features set since it is not in current dev Stockfish. So now, the exe's with the additional features are actually faster with than the exe's without add'l features

Official Stockfish on my machine

Code: Select all

[Mac-Pro:~/cluster.mfb] michaelbyrne% stockfish bench 16 1 16 > /dev/null

Position: 1/46

Position: 2/46

Position: 3/46

Position: 4/46

Position: 5/46

Position: 6/46

Position: 7/46

Position: 8/46

Position: 9/46

Position: 10/46

Position: 11/46

Position: 12/46

Position: 13/46

Position: 14/46

Position: 15/46

Position: 16/46

Position: 17/46

Position: 18/46

Position: 19/46

Position: 20/46

Position: 21/46

Position: 22/46

Position: 23/46

Position: 24/46

Position: 25/46

Position: 26/46

Position: 27/46

Position: 28/46

Position: 29/46

Position: 30/46

Position: 31/46

Position: 32/46

Position: 33/46

Position: 34/46

Position: 35/46

Position: 36/46

Position: 37/46

Position: 38/46

Position: 39/46

Position: 40/46

Position: 41/46

Position: 42/46

Position: 43/46

Position: 44/46

Position: 45/46

Position: 46/46

===========================
Total time (ms) : 10627
Nodes searched  : 19238537
Nodes/second    : 1810345
[Mac-Pro:~/cluster.mfb] michaelbyrne% 

Stockfish-XR7

Code: Select all

[Mac-Pro:~/cluster.mfb] michaelbyrne% Stockfish-XR7 bench 16 1 16 > /dev/null

Position: 1/46

Position: 2/46

Position: 3/46

Position: 4/46

Position: 5/46

Position: 6/46

Position: 7/46

Position: 8/46

Position: 9/46

Position: 10/46

Position: 11/46

Position: 12/46

Position: 13/46

Position: 14/46

Position: 15/46

Position: 16/46

Position: 17/46

Position: 18/46

Position: 19/46

Position: 20/46

Position: 21/46

Position: 22/46

Position: 23/46

Position: 24/46

Position: 25/46

Position: 26/46

Position: 27/46

Position: 28/46

Position: 29/46

Position: 30/46

Position: 31/46

Position: 32/46

Position: 33/46

Position: 34/46

Position: 35/46

Position: 36/46

Position: 37/46

Position: 38/46

Position: 39/46

Position: 40/46

Position: 41/46

Position: 42/46

Position: 43/46

Position: 44/46

Position: 45/46

Position: 46/46

===========================
Total time (ms) : 10396
Nodes searched  : 19238537
Nodes/second    : 1850571
[Mac-Pro:~/cluster.mfb] michaelbyrne% 

Note - fast magics primarily help modern popcount builds, not bmi2 builds ( bmi2 builds do not get fast magics )

Standard bench - official cur-dev-stockfish

===========================
Total time (ms) : 3013
Nodes searched : 5742013
Nodes/second : 1905746
[Mac-Pro:~/cluster.mfb] michaelbyrne%

Stockfish-XR7 ( cur-dev-stockfish with add'l features)
===========================
Total time (ms) : 2934
Nodes searched : 5742013
Nodes/second : 1957059

Both run on modern-popcount processors

to compile Honey without features with the Makefile provided, would be something like this : make profile-build ARCH=x86-64-modern COMP=gcc HONEY=yes FEATURES1=no

to compile Bluefish without features with fortress detection with the Makefile provided, would be something like this : make profile-build ARCH=x86-64-modern COMP=gcc BLUEFISH=yes FORTRESS_DETECT=yes FEATURES1=no

Black Diamond has FT built in regardless , so to build Black Diamond without features : make profile-build ARCH=x86-64-modern COMP=gcc NOIR=yes FEATURES1=no

by default FEATURES1 and STOCKFISH are set to be built, so to build a functionally equivalent stockfish with features, use the standard make command with the Makefile provided : make profile-build ARCH=x86-64-modern COMP=gcc, as an example from my machine.

[PeterO]

Hi Mike,

will you compile the engines for Android too?

Peter

[MikeB]

Hi Mike,

will you compile the engines for Android too?

Peter

We will try , I might need some help. For the Pi definitely.

Fixing "fast magics" seems to have helped, with just 60 games to go, it would be fair to say this will be the strongest version ever.

Code: Select all

440 of 500 games completed...
Time control: 300 seconds + 3.0000 seconds
Threads = 1
Hash = 256
Date: 11/27/19 : 11:44:41
440 game(s) loaded
Rank Name                      Rating   Δ     +    -     #     Σ    Σ%     W    L    D   W%    =%   OppR 
---------------------------------------------------------------------------------------------------------
   1 Honey-XR7                  1056   0.0   15   15   440  223.0  50.7   50   44  346  11.4  78.6  1052 
   2 Stockfish 251119 64 POPC   1052   4.8   15   15   440  217.0  49.3   44   50  346  10.0  78.6  1056 
---------------------------------------------------------------------------------------------------------

With just 60 games, barring some total collapse by Honey this would be one of those rare times where Honey finishes on top against it origin.

Of course, the error bars signify this could be an outlier event, nonetheless , I 'm encouraged by the results that nothing is seriously broken.

Edit: It's now up to 456 games, with the all of the games since the above results ended in draw. Just 44 games to go , about an hour.

[Marek Soszynski]

If I do fix it , will you promise that you will learn how to compile your own ? Rather make this a self service operation.

Thanks Mike. I promise.

Just to explain, the issue isn't about compiling for one's own OS on one's own OS, which in my case is Linux. I do that already. What I will be trying to learn is how to make a Windows compile on a Linux system. (The Windows compiles are for GUIs running under Wine.) Don't wait up.