Author: Andrew Williams
Date: 02:02:20 11/30/00
Go up one level in this thread
On November 29, 2000 at 05:24:20, Andrew Williams wrote: >On November 28, 2000 at 16:06:11, Bruce Moreland wrote: > >>On November 28, 2000 at 14:06:27, Andrew Williams wrote: >> >>>On November 28, 2000 at 12:52:57, Scott Gasch wrote: >>> >>>>Hi, >>>> >>>>I posted a couple of messages about move ordering yesterday and wanted to share >>>>some results from my (limited) testing. I ended up implementing the suggested >>>>"apparently losing captures" (MVV/LVA) after all others order. In one test >>>>position this resulted in a tree 200k nodes larger at 8 ply but in two others it >>>>resulted in a marginally smaller (under 40k nodes) tree at 8 ply. I will do >>>>more testing on this matter but it may be a moot point because I intend to write >>>>a SEE pretty soon. >>>> >>>>I also did some experimenting with ordering captures that take the last moved >>>>enemy piece. At low search depth this seems to make some difference but at >>>>higher depth this heuristic actually grew the tree in all three test positions > I tried. >>>> >>>>I also did some playing with history weight and settled on hist[x][y] += (2 << >>>>depth). >>> >>>I use history[whoseTurn][frsq][tosq] += (depthremaining*depthremaining) >>>I have a separate table for white and black. Every few plies I divide >>>this number back by a lot (can't remember how much or how often). >> >>Why does everyone think that a move that cuts off after a deep search is more >>likely to produce cutoffs elsewhere in the tree, than one that cuts off after a >>shallow search? >> >>That multiplication is expensive. Does it achieve anything? >> >>bruce > >That's a good question. Last night I ran a couple of test sets with my >draft*draft scheme versus a constant increment scheme. I chose the constant >based on what the average of draft*draft works out to be. My draft*draft scheme >solved one more problem in both test sets (ie 412 out of 770 in ECM98 5 seconds >and 56 out of 90 in an old Arasan test suite at 60 seconds). This evening I'll >run a node comparison for a fixed depth. This might provide some results that >are more useful. > >Andrew Last night I ran a test, searching every position in lct2 to depth 10. I measured the total number of nodes, and I got: Draft Squared: 93154844 Constant: 104540199 This is a small test set. I'm trying a larger set at home now. Andrew
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.