Author: Stuart Cracraft
Date: 14:05:38 01/22/06
Go up one level in this thread
On January 20, 2006 at 20:29:04, Joseph Ciarrochi wrote: > > >engines try to exclude a number of bad lines quickly to improve depth of >search.. Let's say one of those bad lines turns out to be good at a depth of 12. >Once the line is excluded, how does the engine go back to it? > >for example, let's say a queen sack looks terrible at depth of 3, and its value >is only discoverable at depth of 10. Once the queen sack is excluded at depth of >3, how does the engine later discover that the queen sack was a good idea? > >naively yours, > >Joseph It discovers it with Iterative Deepening, besides other things. :-) So assume time X for move. You now start an infinite loop where PLY increases from 1 on up to infinity. PLY represents the ply you are searching to from the present real-world board position. So you hit ply equal 3 which is passed to yours search routine as depth=3. That then decrements or increments it as according to reduction or extension for each move in that position. Quiescence is tacked on when depth goes to 0 (or 1 in some cases) and goes out much further. Eventually, your queen sack at ply 1 becomes known to be good because the iterative deepening will be well past ply=3 with the associated quiescence. As each iterative deepening loop iteration completes, the run from the last iteration produces information that greatly improves the next iteration. That information is: sorted movelist at ply 1 is good, more useful hashed positions in the transposition table, killer moves at various ply, and so on. So it is not an expensive approach but a better approach and it helps speed things up for deeper depths. Some have experimented with plies 1-3-5-7-9-etc or searching odd-numbered, but the idea is the same. Earlier searches improve later searches and moves get researched in a different, ostensibly improved order, but not thrown out (in most implementations.) They *used* to throw out moves years ago. But that is too dangerous now. Interestingly, M Chess, nee A.I. Chess, could change its mind back to a previously searched move ON THE SAME PLY, as I recall. There are probably lots of programs that do that these days. If an author with such a feature would speak to that, I think it will help focus in on the specific case the questioner has asked about. Stuart
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.