Author: Omid David
Date: 03:29:41 08/31/02
Go up one level in this thread
On August 31, 2002 at 06:24:18, Bo Persson wrote: >On August 30, 2002 at 18:42:40, Omid David wrote: > >>On August 30, 2002 at 17:39:27, Gian-Carlo Pascutto wrote: >> >>>On August 30, 2002 at 17:37:25, Patrik wrote: >>> >>>>Difference is that Crafty used alpha instead of value when it re-searches. >>>>Is there any reason to use alpha instead of value? >>>>Using value which is greater than alpha seems to cause more cutoffs than using >>>>alpha. >>> >>>If you get a fail high on the first search and a fail low on the second (*), >>>you will lose your PV's. This does not happen if you do the research with >>>alpha. >>> >>>(*) If you think this can't happen, you haven't been doing chessprogramming >>>long enough. >>> >>>-- >>>GCP >> >>If you take a look at Aske Plaat's PhD thesis "Research Re: Search and >>Re-search" and his numerous other publications, you'll notice that on many >>occasions his results are not substanciated enough in practice. >> >>For example he conducts all his experiments (on 20 test positions, depth 8) in >>brute force fixed depth search, which is extinct nowadays (even in 1996, who >>didn't use a form of variable depth search?). >> > >That was done to get numbers that were easy to compare (I think he says so too). Yes. Page 27 of the thesis: "Studying this class of algorithms [fixed-depth full-width] has the advantage that performance improvements are easily measurable; one has only to look at the size of the tree." But as Heinz and many others have shown, variable depth algorithms are comparable too; but in addition to the size of the tree, the tactical strength should also be measured (by experimenting on test suites or self play matches). >If you have extensions and pruning in the tree, *any* changes to the tree will >affects its size. Even a seemingly random event in the search can get an >extension triggered or not. So to have a stable base for his measurements, he >disabled most stuff, and ended up with a fixed depth search. This somewhat >reduced the usefulness, but it *did* produce a result. > > >I have seen that a simple thing like recoding quicksort for the move list can >affect the tree size by 50% in one particular test position. The sorting was >still correct, but moves with the same preliminary score ended up in a different >order. > > > >Bo Persson >bop2@telia.com
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.