Author: Robert Hyatt
Date: 19:52:16 07/26/01
Go up one level in this thread
On July 26, 2001 at 16:32:37, Bruce Moreland wrote: >On July 26, 2001 at 12:41:34, Uri Blass wrote: > >>In the relevant example not trading is a good alternative and there is no >>problem of losing material. >> >>In theory there may be a position when the mistake was 4 plies before the >>trading but practically it does not happen often. >> >>I agree that it is better to solve things by evaluation but it does not mean >>that search does not help. > >Bob is right that there are lots of cases where search does not help enough. We >have the concept of effective depth. An obvious example is doubled pawns. A >program that does not evaluate bad doubled pawns will lose pawns beyond its >search horizon. If it evaluates bad double pawns, it will stop losing these >pawns. Essentially, it is able to avoid having its search depth rendered >ineffective due to big sub-trees that are full of mistakes. > >A fast program with a limited eval function is vulnerable to what I call a "dead >area" in its search. A dead area is a place where all of the evals in a >sub-tree are wrong, and extra depth doesn't make the evals much better. There >are lots of things that can cause vast dead areas, for instance improper >handling of cases involving bishops of opposite color. Extra search does not >help in these cases, so time spent searching is wasted. > >I don't understand why Bob is so hot on this one type of position though. >Crafty didn't always evaluate this kind of pawn structure, and it's not like it >was a bad program before it did. I'm not particularly hot on this one thing. But I _did_ see it get exploited way too many times which caused me to finally fix it. There are lots of other holes that need fixing just as badly of course. Against comps, this is probably not very important at all since most comps don't understand outside passers and outside candidates very well. But all GMs do. And when they notice you don't, trouble starts. Or at least it did for me. > >Perhaps this is an oblique continuation of the Deep Blue/GM argument. Not for me. I don't know what they evaluated here at all, so I can't comment nor speculate. > >There are a lot of cases that you want to catch if you can, but I don't think >that absence of checking for these cases is an a priori argument that a program >sucks. If you told me I could fix any number of these minor cases (pawn >structure, king safety, opposite bishops, insufficient material, etc., are more >essential major cases) that I wanted to fix, and run the same speed I run now, >or I could use my current stuff on a machine twice as fast, *against computers >at least* I would take the faster machine every time. > >bruce I would never say "sucks". I would say "plays horribly in classic pawn endings" however. And I used to fear playing GMs in these positions. I no longer give them a second thought. Crafty doesn't win them all, by any stretch of the imagination. But it doesn't look like a 1200 player and do that BxN capture either, which is (to me) a good thing. :)
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.