Author: J. Wesley Cleveland
Date: 14:02:36 08/19/00
Go up one level in this thread
On August 18, 2000 at 19:06:03, Robert Hyatt wrote: >On August 18, 2000 at 14:48:43, Dave Gomboc wrote: > >>On August 18, 2000 at 13:53:16, J. Wesley Cleveland wrote: >> >>>On August 18, 2000 at 09:23:38, Robert Hyatt wrote: >>> >>>>On August 17, 2000 at 18:05:41, Gian-Carlo Pascutto wrote: >>>> >>>>>On August 17, 2000 at 14:43:08, J. Wesley Cleveland wrote: >>>>> >>>>>>When you have a fail-low in a deep search where the value drops significantly, >>>>>>finding an alternate move can take a very long time. This is largely because >the >>>>>>values in the hash table are largely useless, so in effect we are researching >>>>>>the entire tree. It seems to me one should use iterative deepening, and start >>>>>>from ply 1 again. >>>>> >>>>>This technique has been described by Schaeffer a long time ago... >>>>> >>>>>(Obvious question: Why is nearly no-one doing it?) >>>>> >>>>>-- >>>>>GCP >>>> >>>> >>>>I'm not sure what you mean by described a long time ago. But there is a big >>>>problem. If you start over at 1 ply, you don't get the fail low score. You >>>>find (again) the _wrong_ move, until you get deep enough. When there is a big >>>>score swing between two iterations, you take your lumps. There is no way to >>>>cheat alpha/beta there. >>> >>>You should get the fail low score, since it is in the hash table. They should >>>stay there as they are analyzed to a greater depth than you are likely to get >>>to. >> >>Yes, the technique relies upon the information from the deeper searches being >>present and used to perform cutoffs at shallower searches. >> >>Dave > > >Doesn't this sound _gross_ time-wise? You find another move that doesn't fail >low, until the last iteration when the truth is found. If you repeat this >a few times, it seems worse than just searching for a new best move??? I thought that was what I *was* suggesting. At this point, all you know that move a, which you thought was good, is terrible. So why not treat it as almost a new position and research from ply one. If several lines are all going to fail low at a deep ply, your search is in big trouble anyway.
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.