Author: Robert Hyatt
Date: 14:22:55 10/18/00
Go up one level in this thread
On October 18, 2000 at 15:53:10, Uri Blass wrote: >On October 18, 2000 at 14:42:54, Robert Hyatt wrote: ><snipped> >>The problem is that the 10% wrongs will influence things more than the 90% >>right. only a small fraction of a percent of the total nodes searched affect >>the PV/root score. > >The question is what do you mean by the 10% wrongs. OK. Feed an evaluation function thousands of randomly selected positions. If it produces a wrong answer in 10% of those, then it has a problem, _if_ the search isn't controlled. IE if you are selective, and you simply don't search the parts of the tree where the eval is wrong, or you don't call the eval in positions where it can possibly be wrong, there is no problem. But if the eval is wrong 10% of the time in randomly selected positions, the search will encounter those positions and use them to make poor decisions. If the eval is wrong for often it is even worse, of course. > >I thought that the meaning is 10% of the positions when there is big positional >scores and not 10% of the positions in the tree. You could take it either way. I really mean that if a term is used to produce a big score, and that term is wrong 1 of every 10 times, then "interesting" things are going to happen. From experience. And note interesting != good, in this case... > >If the meaning is 10% of the positions in the tree than it is different because >it means that in most of the cases that there is big positional scores the big >scores are not justified. > >If the meaning is 10% of the position when there is a big positional score than >I assume that other programs have more problems from not having big positional >score in 90% of these cases > >Uri I simply worry about the cases where a specific term is wrong even 10% of the time. The search can use that against you in ways you would never dream of, due to the full-width stuff.
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.