McBrain v2.1 Release

[MikeB]

All details and flavors available here:

https://github.com/MichaelB7/Stockfish/ ... /tag/v2.1a

Enjoy!

ps: this is a 2 for 1 deal, you get two versions , v2.1 and 2.1a.

[MikeB]

some may find this interesting
test results using the 40,000 + 2 ply positions that fishtest uses:

Code: Select all

1000 game(s) loaded
Rank Name                      Rating   Δ     +    -     #     Σ    Σ%     W    L    D   W%    =%   OppR 
---------------------------------------------------------------------------------------------------------
   1 Stockfish 010417 64 POPC   3127   0.0   17   17   400  219.0  54.8   92   54  254  23.0  63.5  3093 
   2 McBrain 2.1s 64 POPCNT     3121   6.0   17   17   400  214.5  53.6   86   57  257  21.5  64.2  3095 
   3 McBrain 2.0 64 POPCNT      3121   0.2   17   17   400  214.0  53.5   95   67  238  23.8  59.5  3095 
   4 McBrain 2.1f 64 POPCNT     3115   6.0   17   17   400  209.5  52.4   93   74  233  23.2  58.2  3096 
   5 Komodo 10.4 64-bit         3016  98.6   17   17   400  143.0  35.8   57  171  172  14.2  43.0  3121 
---------------------------------------------------------------------------------------------------------

note : all positions were chosen at random for the games above from the SF test file below ( I believe their original file may have been in pgn format, I converted the pgn format to epd - it's been a while, so my recollection is hazy).

https://www.dropbox.com/s/4qetc5mg8du1v ... d.zip?dl=1

to say these positions are "offbeat' would be an understatement - you can check for yourself..

below, same engines, same time control, same hardware - but using selected 8 ply openings played by ICCF Grandmasters:

Code: Select all

1000 game(s) loaded
Rank Name                      Rating   Δ     +    -     #     Σ    Σ%     W    L    D   W%    =%   OppR 
---------------------------------------------------------------------------------------------------------
   1 McBrain 2.0 64 POPCNT      3126   0.0   16   16   400  218.5  54.6   85   48  267  21.2  66.8  3093 
   2 McBrain 2.1f 64 POPCNT     3121   5.3   15   15   400  215.0  53.8   75   45  280  18.8  70.0  3095 
   3 McBrain 2.1s 64 POPCNT     3119   2.5   16   16   400  212.5  53.1   76   51  273  19.0  68.2  3095 
   4 Stockfish 010417 64 POPC   3096  22.8   16   16   400  196.5  49.1   75   82  243  18.8  60.8  3101 
   5 Komodo 10.4 64-bit         3038  57.3   16   16   400  157.5  39.4   57  142  201  14.2  50.2  3115 
---------------------------------------------------------------------------------------------------------

It's not surprising that K plays so much "better" here , since I am certain the K team are not using SF 40,000 2 ply opening for tuning , all of the SF engines here have exposure to that type of opening tuning, with the current SF engine having the most exposure , while the McB engines are not being tuned with the SF openings at all...

the time time control for all these games were the standard SF 10 sec with 0.1 inc, which is 9.5 sec with 0.1 inc on my machine when adjusted by fishtest.

[Henrik Dinesen]

Thanks Michael

In your tests you show the s-version (apart from f and 2.0), while in your initial post you refer to the a-version ... A secret reason, delayed Aprils-fool or a typo perhaps ?!

Best,
Henrik

[MikeB]

Thanks Michael

In your tests you show the s-version (apart from f and 2.0), while in your initial post you refer to the a-version ... A secret reason, delayed Aprils-fool or a typo perhaps ?!

Best,
Henrik

Typo - for test purposes , it was 21a, 21b 21c etc up to 21s. Version 21.f became v21.1 and 21.s became v2.1a, I should have fixed that to make it
more clear.

[Henrik Dinesen]

Ok, it's clear now. I was watching for this release

[Paloma]

> Version 21.f became v21.1 and 21.s became v2.1a,

v21.1 ? Another typo? I see no 21.1 in your DL

Nevertheless thanks for providing this

[MikeB]

> Version 21.f became v21.1 and 21.s became v2.1a,

v21.1 ? Another typo? I see no 21.1 in your DL

Nevertheless thanks for providing this

yes - Was rushing before heading out 2.1 and 2.1a is correct

[MikeB]

All details and flavors available here:

https://github.com/MichaelB7/Stockfish/ ... /tag/v2.1a

Enjoy!

ps: this is a 2 for 1 deal, you get two versions , v2.1 and 2.1a.

Further testing revealed (12000 games) with "FindMate" set to true (default) , McB 2.1a is about 20 ELO weaker than current SF , with it set to false , it's virtually the same ELO as SF, no real surprise there. Going forward , the default will be set to false. Tc 30 sec with 1/2 sec increment

[Vinvin]

All details and flavors available here:

https://github.com/MichaelB7/Stockfish/ ... /tag/v2.1a

Enjoy!

ps: this is a 2 for 1 deal, you get two versions , v2.1 and 2.1a.

Further testing revealed (12000 games) with "FindMate" set to true (default) , McB 2.1a is about 20 ELO weaker than current SF , with it set to false , it's virtually the same ELO as SF, no real surprise there. Going forward , the default will be set to false. Tc 30 sec with 1/2 sec increment

For me 20 Elo weaker than SF is not very important : analyze with an engine rated 3300 or 3280 is not a big difference. I prefere to get an engine that gives :
1) A better evaluation e.g. "mate in 25" instead of "+35.00"
2) An engine that gives an analyze more accurate. Sometimes SF gives wrong lines for specific positions (zugzwang, fortress, some endgames, ...)

[Nordlandia]

Code: Select all

For me 20 Elo weaker than SF is not very important : analyze with an engine rated 3300 or 3280 is not a big difference. I prefere to get an engine that gives : 
1) A better evaluation e.g. "mate in 25" instead of "+35.00" 
2) An engine that gives an analyze more accurate. Sometimes SF gives wrong lines for specific positions (zugzwang, fortress, some endgames, ...)

McBrain with Nalimov support will be fantastic addition.

Something on the implementation agenda`?

I find DTM format useful for Let's Check analyse and endgame studies.