Author: Robert Hyatt
Date: 14:56:51 01/20/01
Go up one level in this thread
On January 20, 2001 at 16:54:33, Vincent Diepeveen wrote: > > >On January 19, 2001 at 21:27:52, Uri Blass wrote: > >>On January 19, 2001 at 19:44:49, Vincent Diepeveen wrote: >> >>>On January 19, 2001 at 17:23:03, Scott Gasch wrote: >>> >>>>Hi all, >>>> >>>>Here's a pretty simple test position. This comes from the WAC suite... my >>>>engine does fairly well on this suite as a whole but seems blind to this >>>>solution. >>>> >>>>[D]8/7p/5k2/5p2/p1p2P2/Pr1pPK2/1P1R3P/8 b - - >>>> >>>>The solution is Rxb2 -- black's connected passers are unstoppable after the >>>>recapture. >>>> >>>>My program refuses to find this solution... even at 9 ply it misses it. The >>>>strange thing is that from the other side after Rxb2 it sees that white is toast >>>>very quickly... score dropping to -500 or so after about 1 second. >>>> >>>>My question is, of course, how this move is missed. I've tried kicking up the >>>>value of connected passers and passed pawns in general. I've tried adding a >>>>special rule to eval about connected passers on the 7th, on move, with control >>>>of a queening square. I've tried cutting back my futility margin in qsearch and >>>>always extending a full ply for checks (It usually extends only 3/4 ply for >>>>checks after the iteration depth). And still it does not find Rxb2. >>>> >>>>Even stranger is if I run a static eval with the two connected passers rolling >>>>towards the queening square after the rook exchange the eval puts black ahead! >>>>I can't seem to figure this out... either my pruning is too aggressive or there >>>>is some other bug in the engine...? >>>> >>>>I hope someone out there can give me a little advice. Thanks! >>>> >>>>Scott >>> >>>First of all this combination is not 9 ply but something >>>like 14 ply. >> >>The number of plies is dependent on the evaluation function and on the >>extensions and pruning rules. >> >>If the evaluation function knows that the following structure is good for black >>then you do not need a lot of plies >> >>[D]8/8/5k2/8/8/2ppPK2/1R6/8 w - - 0 1 >> >>The white pawn at e3 is of course important in this structure because it blocks >>the white king. >> >>The program should know that without the white pawn at e3 the structure is good >>for white. >> >> >> Most programs solve it incorrectly at small depths as >>>they don't evaluate kings stopping passers or because they just >>>give huge scores for doubled passers. >> >>I think that most programs do not solve it at small depthes. >> >>Uri > >You are right about the e3 pawn. This is the ONLY reason why >it wins for black. See the evaluation for many programs of this position >without the pawn on e3 and you'll see they statically evaluate it as +2.0 >for black or something similar. > >+3.0 for each passer, 1 pawn more minus rook is +2.0 to roughly >do the math. > >Of course search depth depends upon what you do in qsearch, >selective search, whether you extend all checks or not >and whether you have passed pawn extensions not to mention >threat extensions. > >So nowadays i solve this at i think 10 ply or something because of >the extensions, but i only solve it tactically; so i see >that i get a queen with black and not evaluation is solving it. > >What i find wrong are programs that solve this position positionally. >No evaluation takes into account the e3 pawn i'm 100% sure of that. >I don't do it in DIEP either. Because what if it's on e2 and someone >attacks e3? > >There are zillions of possibilities, but a prog simply shouldn't >solve this without seeing it the pawns promote to a queen! You should look at the game last night between crafty and diep before you say that... If you wait until you see a promotion, you are going to get killed. I'd rather say two connected passers on the 6th win and be right 90% of the time, rather than not saying that and being wrong 90% of the time. On the other hand, you are evaluating isolated pawns at a ridiculously high rate. I watched a game where you had pawns on the gh files (2) and Crafty had pawns on the f and h files (3). You thought you were over a pawn ahead, and got zapped for thinking that. I don't think _any_ evaluation is right 100% of the time. I am pretty convinced mine is right more often than it is wrong, and getting more accurate all the time. > >Even worse as this is the gs2930 testset, where programs can >give away pawns for nothing! I don't give away pawns for _nothing_. :) > >Greetings, >Vincent
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.