Author: Peter Fendrich
Date: 15:17:24 02/07/99
Go up one level in this thread
On February 07, 1999 at 01:43:40, David Eppstein wrote: >I know that the normal aspiration search windows are usually something like >previous search value +/- delta, where delta is a fixed parameter e.g. 1/2 pawn. > >But has anyone tried setting the window in a more complicated way that >incorporates information from the last several searches? One could get an >estimate of the variance of scores from search to search, which might help >narrow or widen the window in situations where that helps. For instance, maybe >in place of delta one could use (some parameter times) the difference between >the last two search scores. It's might also be possible that a non-centered >window would sometimes be appropriate. > >I know Bob and others have collected statistics on how often the best move >changes in deep searches; are there similar statistics on how much the eval >changes and whether there's much correlation between successive changes? Like you, I think that things can be better done here than just a fixed value. What I do is not so complicated: 1) Sometimes each iteration gives about the same eval, and even the same move. When (counted from iteration 3 and forwards) this happens I narrow the window bit by bit. 2) In some cases there are two moves competing for the first place. One move wins when depth is odd and the other move when depth is even. If this is the case and the first move in next iteration fails low, I just continue with the second move to find out if it will change again. The second case is under testing and re-design. The first move is not always failing low and the second move gets a better eval anyway. I will try some other variations, like deliberatly search the second move first. //Peter
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.