Author: Robert Hyatt
Date: 16:30:40 08/22/05
Go up one level in this thread
On August 22, 2005 at 18:49:25, Uri Blass wrote: >On August 22, 2005 at 17:40:38, Robert Hyatt wrote: > >>On August 22, 2005 at 15:50:15, David Dahlem wrote: >> >>>What is the most likely cause of a huge node explosion at depth 10, with a >>>branching factor of 24? >>> >>>I'm getting 72% move ordering with this order... >>> >>>1. Best move in previous ply >>>2. Winning captures >>>3. Killer moves >>>4. Neutral captures >>>5. Bad captures >>>6. Rest of the moves >>> >>>Thanks for any clues >>>Dave >> >> >>A pathological case. >> >>For example, you search to depth N, and you get all the move ordering stuff set >>for that tree, then you search one ply deeper and discover everything changes, >>and what was a good move in the last search is now horrible because you are >>seeing some tactics you were overlooking at one ply less. >> >>This doesn't happen that often, but it does happen, and there's nothing you can >>do since you can't tell that ordering is screwed up until after the fact... :) > > >If you detect some pattern you can probably do something. > >If your hash table move is move X and you find that X does not fail high but Y >fails high and you find that pattern again and again at the same ply then it may >be better not to start from the hash move but to start from the killer move. > >I did not try it but I do not accept the claim that there is nothing that can be >done(I did not try to do something about it but it does not mean that nothing >can be done). > >Uri Suppose that for one move things falls apart, but not for every move. The problem here is that we don't know that things have gone bad until after they have done so... This is like the blackjack morons that hawk "betting progression systems to exploit patterns in the game of blackjack." None of them work. Because you can't recognize a pattern until it is over. And by then it is too late to change your bets and make more money... I won't say it is "impossible" to fix the pathological search problems. But it is effectively impossible since they don't happen enough to warrant years of research to look for a solution that might or might not exist...
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.