ShashChess learning tested

[Rodolfo Leoni]

Hi. After some years I'm here again with two tests I ran on ShashChess learning system.

First one: 100 games match vs a version of SF NNUE, TC 5+3, on the position

r1b1kb1r/2q2p1p/p2p1P2/2nP4/1p1N4/5Q2/PPP4P/2KR1BR1 b kq - 0 18

This is the result:

-----------------SF_NNUE_x64_240720_AllAVX_POPCNT_256_msys2-----------------
SF_NNUE_x64_240720_AllAVX_POPCNT_256_msys2 - ShashChess13.1-x86-64-modern : 40,0/100 12-32-56 (101===1=10=0===01====0===0=0=0=0===0==1========0====1==0=0==0==0=0===0=01010=0=0=0=0=010=01010=0=0=0) 40% -70
-----------------ShashChess13.1-x86-64-modern-----------------
ShashChess13.1-x86-64-modern - SF_NNUE_x64_240720_AllAVX_POPCNT_256_msys2 : 60,0/100 32-12-56 (010===0=01=1===10====1===1=1=1=1===1==0========1====0==1=1==1==1=1===1=10101=1=1=1=1=101=10101=1=1=1) 60% +70

Note that after 25 games performance was 50% and learning had a visible effect from game 26 to end.

--------------------------------------------------------------------------------------------------

Second test:

Learning file was built from ground by game analysis of games of first test at fixed depth 25. Then I repeated the 100 games match with the "read only" UCI option turned on and I got these results:

-----------------SF_NNUE_x64_240720_AllAVX_POPCNT_256_msys2-----------------
SF_NNUE_x64_240720_AllAVX_POPCNT_256_msys2 - ShashChess13.1-x86-64-modern : 33,0/100 17-51-32 (=010=0=0=010=000=000=0=0=0=0=01==0=010=0=0=0101010101010=0=0=0=01010=0=0=010=010=0=010=01010=0=0=0=0) 33% -123
-----------------ShashChess13.1-x86-64-modern-----------------
ShashChess13.1-x86-64-modern - SF_NNUE_x64_240720_AllAVX_POPCNT_256_msys2 : 67,0/100 51-17-32 (=101=1=1=101=111=111=1=1=1=1=10==1=101=1=1=1010101010101=1=1=1=10101=1=1=101=101=1=101=10101=1=1=1=1) 67% +123

---------------------------------------------------------------------------------------------------

Conclusions:
The system works fine. It could be a bit more user-friendly if one could decide the learning file name instead of having the experience.bin file, so that he can build a specific learn file for his favorite openings and positions, naming them.

Best wishes,
Rodolfo

[carldaman]

Unless the experience file grows quickly and dramatically, I would be more than happy with a single exp file for everything.
Some openings can transpose into other openings.