Author: Robert Hyatt
Date: 20:05:30 02/02/04
Go up one level in this thread
On February 02, 2004 at 14:41:22, Uri Blass wrote: >On February 02, 2004 at 10:15:31, Robert Hyatt wrote: > >>On February 01, 2004 at 18:31:20, Ed Trice wrote: >> >>>Hello Dr. Hyatt, >>> >>>>On February 01, 2004 at 17:56:18, Ed Trice wrote: >>>> >>>>>I am wondering if Fritz has a specific evaluation routine for indentifying drawn >>>>>Rook + f + h pawns vs. Rook arrangements. In the Crafty-Hiarcs game from the >>>>>last round of the tournament, it was kibitzing a draw score for quite some time. >>>>> >>>>>Could it be possible to encode such a large subset of knowledge into a faultess >>>>>"tablebase equivalent"? Or was it making the correct prediction for an incorrect >>>>>reason? >>>> >>>> >>>>Fritz was wrong. It kibitzed 0.0 for 50 moves. But if you remove rooks, white >>>>wins instantly due to distant majority/passer. Fritz 7 did not have a clue >>>>about that... >>> >>>Thanks, I came in around move 45 or so, I did not see the earlier announcements. >>> >>>Given that, I am wondering to what extent some of the R+h+f vs. R knowledge can >>>be encoded in a way similar to (Heinz 2003) from Advances in Computer Games 10. >> >>It is not hard. I have some code that seems to work, although it is not in the >>current version. It fits right in with the current KRP vs KR evaluation code >>that is present. It probably needs a bit more work, just like the recent >>changes to the KR[BN] vs KR stuff that had an exception I had missed. R/B pawns >>is not that easy to draw, so you have to catch anything that breaks it. I wrote >>some test code to run thru a bunch of KR[BN] vs KR positions and summed up wins >>and draws, to see where the problems are. It is not very easy to actually code >>things up, although king on the edge is pretty good. But it is not perfect, as >>a game we played while working on the book (against Hiarcs) showed. We were >>winning, but we went into a KRB vs KRPP (we had KRB) and thought we were winning >>because Hiarcs had the king on the h-file with our rook on the g-file. But we >>could not ever win that game... The exceptions are a pain. :) > >You cannot know by evaluation if it is a win or a draw otherwise you do not need >search. > >computer does not "think" in terms of winning but in term of score and if you >have something like +1 score it does not mean winning but something like >probability of 50% to win and 50% to draw. > >If you think that your evaluation can cause you to miss a win too much by going >to KRB vs KRPP than divide the evaluation by 2 if the side with the bishop has >the advantage. > >Uri I am not sure what you are saying, but it _is_ possible to evaluate certain classes of endings as won, and other similar positions as "unknown." I do this in several places. IE take the wrong-bishop + rp positions. Some you can say "draw" some you can't. You let the search try to reach draw or no draw as best it can, so that the evaluation can say for sure. Ditto for KR +BP +RP vs KR. It is the same kind of problem. It's a trivial win if the king is behind the pawns too far, it is a difficult draw if the king is in front... The divide by two is _not_ enough. KRB vs KRP would _still_ be +1, but it is a dead draw in most cases. There are certain types of positions you simply have to say "drawn" or "not drawn" as reliably as possible..
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.