Author: Uri Blass
Date: 09:18:59 09/28/02
Go up one level in this thread
On September 28, 2002 at 11:38:19, Sune Fischer wrote: >Carlos Pagador just sent me a game where frenzee made a clear blunder move. > >It wasn't a bug in the search, but in the time management. >What happened was that the pv move failed-low at the root (I think that's what >it's called?), it was a mate in 3 so it had to be avoided. > >The searched continued and the second move searched happened to be a very bad >queen "sacrifice"! > >Unfortunately time was up before it could search the third move, so it played >this losing queen move. > >It could have been worse actually, if it hadn't searched the second move either >it would have gone straight into the mate, not suspecting the move was bad at >all. > >I wonder how many buggy moves are made because of these fail-lows, I never >thought about this at all, but of course the actual move being returned could be >almost random when this happens. I reckon this is common knowledge, I just don't >remember having seen it explained anywhere? It is a common knowledge and Amir Ban explained in CCC some years ago that Junior almost always tries to finish the iteration that it is searching and only in cases when there is a danger of losing in time it does not stop in the end of the iteration. I also do it and my decision if to start a new iteration is based on my estimate of the time to finish the next iteration. There are programs that do not do it but every good program that I know use more time after failing low to try to finish the iteration. I know it based on my experience in watching games of chess programs togrther with their evaluations. Hiarcs even extend time after finishing the iteration and it may be too dangerous but the idea is that if the position failed low the program probably does not understand it and it may be important to search another ply to avoid the horizon effect. I also know about other mistakes that they do in time management and movei today also does part of their stupid mistakes. Here is one small example: Suppose that your program is at move 41 and the time control is 40 moves for 40 minutes. Suppose that your program wants to play a move that force a draw by repetition A smart time management is not to play it but to try to find alternative that is clearly better and to play the drawing move only when there is danger of losing on time because you can lose nothing by delaying the decision. I prefer not to work on fixing it because the 1 elo that I may gain from it is not worth the danger of new bugs. > >Ok, the fix is obvious, namely to finish the ply if your first move "fails-low", >and somehow ask for a time extension to make it happen. Programs usually do it. Uri
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.