Author: milix
Date: 01:14:51 01/24/05
Go up one level in this thread
Hi rasjid The problem of chess we try to solve is to play the best move in a given position. If I had a perfect evaluation function, I need no search at all. I can tell you the best move without searching. But since perfect evaluation is not possible we need to search as deep as we can to be sure that our evaluation is reliable. Offen deeper searches results in better evaluation. Extensions are a way to get better evaluation and avoid horizon problems - that causes bad evaluation. I think we dissagree in terminology, not the concept! Checks take away plies by forcing a side to get out of check. So checks must be extended everyware (well i do some trick to not extend every check when the side that gives check is winning anyway). But if we have alot of depth to evaluate correctly a passed pawns then why to extend the push in 7th rank? On January 23, 2005 at 18:29:42, rasjid chan wrote: >On January 23, 2005 at 13:41:20, milix wrote: > >>On January 23, 2005 at 11:37:51, rasjid chan wrote: >> >>Hi Alvaro and Rasjid! >>Why search potentially better deeper? I think extensions apply in those >>situations where the value of a line is not reliable enough to be trusted - like >>a line full of checks or recaptures for example. There are some moves that are >>very bad and we need to search really deep to see that they are bad. As for the >>initial question i think that some kind of extensions is beneficial to >>enable/disable them if we are near or far from the leaves. Why to extend a push >>of a passer in the 7th when you are planning to do a 20ply search from there? >> > >As for the need to extend to confirm certain lines are bad or threatening, >I have not come to it yet; but I have read about mate threat extention. > >My earlier statement about searching good moves deeper is inexact. The probably >correct criterion is to search moves that have a higher probability of being >a PV move deeper. > >Chess programming is all about getting a PV that is deeper than what our >opponent can search. Even if we just do a material searchh, we can still win >Shredder provided at every turn we have a PV that is much deeper than what >Shredder can reach plus,say, a factor that Shredder has a good evaluator >(even the best evaluation may not be able to "see" 3/4 plies ahead). > >This being the case(?), then it is irrelevant how far we are from the root >or leaves. If we have a passer at 7th and we search to a depth of >20 , we still extend. The question is whether this passer is the PV. >The PV changes all the time depending on how deep we search. If we can search >to a deoth of 20, the other side probably can match us in depth. So >it comes back to whether this passer is the PV and only searching as deep as we >can can we know the answer. > >Caveat - Errorneous reasoning and fallacies are not intentional. >Rasjid > > > > > > > > > > > > > > > > > > > > > > > >>>On January 23, 2005 at 06:09:23, Alvaro Jose Povoa Cardoso wrote: >>> >>>>Hi everyone, >>>>an issue that has been on my mind for quite some time is that if extensions >>>>should be done everywhere in the tree or at the nodes immediately before >>>>qsearch. >>>>Take for example the pawn push to 7th (2nd) rank extension. >>>>If we are near at the root (say ply2, ply3, ply4, ...) why extend if we are sure >>>>the search will reach and past well beyond the promotion? >>>>Woudn't it make more sense to make this (and other extensions) at nodes >>>>immediately before qsearch ? >>>>Same thing for check extensions, one reply extension, recapture extension, etc. >>>> >>>>Best regards, >>>>Alvaro Cardoso >>> >>>Why not view all extentions this way. >>> >>>We have different lines of play (moves) and we extend lines that seem to have a >>>better chance to be good(PV), ie the likely better moves are searched deeper >>>RELATIVE to the average moves. So the question of your concern may be >>>irrelevant. At the next iterration at root, your good lines will again be >>>searched N depth deeper... repeatedly. This may correct way to view depth >>>extention. >>> >>>Rasjid
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.