Author: Stuart Cracraft
Date: 19:08:23 09/05/04
Go up one level in this thread
On September 05, 2004 at 20:33:45, Michael Henderson wrote: >On September 05, 2004 at 20:08:07, Stuart Cracraft wrote: > >>Odd improvement... >> >>I had used see within quiescence to discard moves >>< 0 of see value (losing combinations) and those >>that didn't make it past a certain margin (futility >>cutoff / delta pruning). >> >>At the time I had added a capability in the move >>generator when a move is added that scored any >>capture with see and stored it for later use. >>This was then used in the quiescence as above >>for discarding, avoiding using the generate there. >> >>But because of having to do the see for each move >>in the main search, the net result was worse performance >>on test suite, not better. >> >>So I just made a separate routine to keep the see() running >>completely out of the main search, do it for all capture >>moves all the time in quiescence but at time move is added >>to movelist instead of at time move is searched. This resulted >>in 1% improvement in score: >> >>wacnew >> >>**** 6.80/23.83 80% 240/300 250.67 73082144 243607/1/291553 0/871282/397085/1606 >>781/18010090/63447 >> >>wacnew with the see() call in movelist routine instead of quiescence, >>but still used in quiescence >> >>**** 6.81/27.35 81% 244/300 249.71 74112552 247042/1/296798 0/854177/1447453/515 >>821/17298092/99799 >> >>Not quite sure why this occurred. There may be some other confounding factor >>that I just got lucky with when making the change. >> >>I looked more carefully and saw a side change I had made was simply to add >>the "see" score into the mvv/lva and then make that the pre-search score >>of the move from the move-generation routine. >> >>Previously, mvv/lva was used only, unless the global see() use in move >>generation for all captures for both main search and quiescence search was >>used. >> >>Now, see score is added in, with mvv/lva, in move generation, for all captures >>but only in quiescence -- see is not used in main search. >> >>Anyway, add your see score in like the above and see if you experience >>a small jump, rather than just using <0 to decide whether to cut. You >>get the benefit of the sort, helping the move ordering. >> >>I speak only to the ones who didn't do this already. >> >>Stuart > >Are you doing 1 second searches? Yes. And before some recent bugfixes in the search, SEE hurt 1 second searches. But since the fixes, SEE helps the 1 second searches. I use a qa program that runs a 6 hour set of all the possible permutations of the best conditional compilation flags. SEE routinely boosts the score.
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.