Just as a reminder: Louis Zulli found some better setting for an 8core system (see here http://www.talkchess.com/forum/viewtopi ... 16&t=31388), and after the release of 2.1, I thought why not give it another try.
My first test is to find the optimal setting for MNTpSP. Stockfish bench command (stockfish bench 512 4 22) over 50! different opening positions, 5 runs, 512 MB Hash, depth 22. MNTpSP from 2 to 8, MSD=4 (fix)
My system: Q6600@3,16GHz, 8 GB Ram, OS LinuxMint 10 64bit
Code: Select all
MNTpSP/MSD 2_4 3_4 4_4 5_4 6_4 7_4 8_4
sec 922 895 1089 962 1018 888 1034
nodes/sec 3732950 3828857 3878249 3869189 3880205 3867726 3868531
957 951 1013 936 1038 1153 899
3724098 3842708 3872926 3865703 3885754 3895055 3865837
884 1024 1037 1015 965 1079 1075
3716748 3838886 3869848 3884031 3869045 3879746 3872916
982 1052 1055 904 1004 957 1033
3740794 3858635 3879112 3868528 3891238 3883996 3878028
891 979 945 940 996 1073 998
3717818 3836214 3862928 3868798 3879267 3879151 3875439
------------------------------------------------------------------------------------------------------------------
Average 927 980 1029 951 1004 1030 1008
3726482 3841060 3872613 3871250 3881102 3881135 3872150
Second best is MNTpSP=5.
Should I refine the results by doing some 2 or 3 additional runs?
My next test will try to determine the best value for MSD.
Same testing scheme, with 50 opening positions and 50 endgame positions. Of course, with a different search depth for the endgame positions.
But which value(s) for MNTpSP should I test? Which one would you choose? Only 5 or 2 as well? Or 3 and 5?
Next question is what are reasonable numbers for MSD?
Which leads me to another question.
In thread.cpp line 112, MSD is muliplied with ONE_PLY. So I get MSD=10 when I set MSD=5 in uci option. Is that right? (Well, multiplying with 1 wouldn't make much sense here, right?)
Hence we only get even nunbers. Is there a reason for doing this? Maybe MSD=7 or 9 would do better. Who knows?
Any advices/comments are welcome.