Author: Omid David Tabibi
Date: 14:52:50 11/22/02
Go up one level in this thread
On November 22, 2002 at 17:06:58, Sune Fischer wrote: >On November 22, 2002 at 12:21:34, Omid David Tabibi wrote: > >>On November 22, 2002 at 12:08:12, Sune Fischer wrote: >> >>>On November 22, 2002 at 07:00:01, Omid David Tabibi wrote: >>> >>>I think you are right, search times are no good, for many reasons. >>> >>>However, why don't you use nodes to solution, rather than nodes to depth? >>> >>>The priority is to solve the position as fast as possible, nodes to solution is >>>a direct measure of that. >>> >>>If you measure nodes to ply 10, what does that say? >>>It doesn't say a lot, I can get to ply 10 in 124 nodes, but the program won't be >>>any good. So you need confirmation that you didn't wreck it by running the test >>>suite. >>> >>>Instead of having the test suite be an indirect verification test, why not use >>>it directly? >>> >> >>Nodes to solution is a great idea. But there are some positions that need a >>tremendous amount of time to be solved. >> >>That idea will be practical only if we have a pool of positions that can be >>solved within a reasonable time. > >That is true. > >But what do you conclude if a new algorithm produces a smaller tree, but also >solves fewer positions, or vice versa; Solves more positions but also produce a >larger tree? > Then I can't conclude anything. But for this paper, I wanted to outperform standard R=2, which is widely known to be better than standard R=3 (see Heinz 1999). The presented results show that the tree size is smaller than std R=2, and the tactical strength greater. That's pretty much enough to show vrfd R=3 is better than std R=2. >I think that must be the danger of aiming for two seperate optimizations. > >-S.
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.