Author: Sune Fischer
Date: 15:54:28 02/06/02
Go up one level in this thread
On February 06, 2002 at 18:23:22, Dann Corbit wrote: >On February 06, 2002 at 18:03:41, Sune Fischer wrote: >[snip] >>Suppose you have millions of given random test positions in which you *know* the >>best move(s). >>Now run tests to see how often a 1-ply search will find the correct move, and >>how often a 2-ply search will find the correct move etc. >>Line up all these percentiles, and you will probably get something like this: >> >>1-ply search: 40% correct moves >>2-ply ......: 55% correct moves >>3-ply ......: 65% correct moves >>4-ply ......: 72% correct moves >>etc... >> >>The thing is, that the percentiles _must_ converge towards 100, so it will need >>to slow down, there may only be 2% difference between a 12 and 13 ply search, >>which is why it is really hard to measure anything. > >Unless it is a forced checkmate, I do not believe any evaluation that says some >move is the best. I only think it might be. I'm not saying that it would be easy to get all those test positions, but if you had them the test would be nice. I suppose one could do a 14 ply search as an approximation to the best move, it wouldn't change the distribution all that much I think. -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.