Author: Bas Hamstra
Date: 01:55:24 09/20/99
Hi, I experimented somewhat with movesorting, what is used as best move and what not and what to store in the HT. You get different results on different positions, what is better in pos1 is worse in pos2. So I ran 20 random wac positions and counted the total nodes. Based on that, I think it is *not* a good idea to keep old best moves (in stead of overwriting) in case the new move is a lowerbound one. This strategy worked badly... So it seems better to not use BM's at all than old BM's from a shallower depth. (remark: my hash replacement scheme only looks at depth). I also tried to *not* store upper records at all, not even the values. On some positions it did quite well but overall it's not so good. My current scheme (which works best so far) is: - Store everything (the normal upper/lower/exact) - If the retrieved hashmove is a upperbound type move simply wipe and ignore it What I didn't try is keeping old best moves (only in same position of course) in case of UPPER stores *if* and only if they have the right depth. I think that might very well be optimal. Realized it too late, no time to complete this test. Fiddling with this can mean a lot to treesize, it seems. I've seen 25% difference on some positions after only minor tweaks. Regards, Bas Hamstra.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.