https://github.com/MichaelB7/Stockfish/ ... /tag/v2.4a
One patch from the SF team that looked really good in testing:
https://github.com/official-stockfish/S ... /pull/1125
Another patch was a from a failed patch back in October which I modified to work with current stockfish code
http://tests.stockfishchess.org/tests/v ... 10626b968c
The patch was designed to continuously tracking of history-mean-value to better calculate rHist. One issue was that it was tracking history-mean-value every 256 nodes which is very expensive. I changed it to about every ~2M nodes. Naturally, there would be no point in submitting such a patch to fishtest since the code would hardly ever be triggered at the tc they use for testing. I ran about 800 games at 5/5 , liked what I saw and I hope you do too. 800 games is not sufficient to say one way or the other , I just liked what I saw. ( I did perform test at 64K, 256K, 512K , 1024K and it was only at 2048K that I saw some benefit - maybe). There was no apparent penalty at very fast tc using the new code that I saw.
This gets rid of the "correction factor" ( see below) which I never quite understood before, but the code for the failed test shows what it was correcting for - not that I understand every little nuance.
Code: Select all
ss->history = cmh[moved_piece][to_sq(move)]
+ fmh[moved_piece][to_sq(move)]
+ fm2[moved_piece][to_sq(move)]
+ thisThread->history.get(~pos.side_to_move(), move)
- 4000; // Correction factor( although you can still grab the source for v2.4 if you wish).
If there was ever a version to test at LTC with multiple cores, this is the version - I hope some of you do that and let me know what the results are and what you think.
As always, all flavors are available for various OS.
Please enjoy!
