Author: Robert Hyatt
Date: 06:48:55 07/07/01
Go up one level in this thread
On July 07, 2001 at 03:41:47, Tony Werten wrote: >On July 06, 2001 at 23:08:07, Robert Hyatt wrote: > >>On July 06, 2001 at 14:27:39, Artem Pyatakov wrote: >> >>>I was taking a peek at Crafty's source code and came across this comment inside >>>the Search() function: >>>---------------------------------------------------------- >>>| | >>>| if there is no best move from the hash table, and this | >>>| is a PV node, then we need a good move to search | >>>| first. while killers and history moves are good, they | >>>| are not "good enough". the simplest action is to try | >>>| a shallow search (depth-2) to get a move. note that | >>>| when we call Search() with depth-2, it, too, will | >>>| not have a hash move, and will therefore recursively | >>>| continue this process, hence the name "internal | >>>| iterative deepening." | >>>| | >>> ---------------------------------------------------------- >>> >>>The comment makes complete sense and "Internal Iterative Deepening" sounds like >>>a great idea, but could somebody please explain to me how to integrate this >>>algorithm with the Killer and/or History heuristics? Thanks in advance. >> >> >>Simple. If this is a PV node, use IID. If it is not a PV node, then use >>the normal history/killer stuff as that is good enough. But on the PV, you >>are searching one ply deeper, and the history/killer stuff is not as accurate. > >Hi Bob, > >from testing I found out it saves a little less than 2%. Should it be more or is >it ones of these things that don't hurt except in a couple of positions where >not having it would kill you ? > >cheers, > >Tony What I find is that it often saves nothing, because it is not done. But on certain types of positions, such as where you fail high at depth N, but the re-search fails low so you get no real score, the depth N+1 search is greatly accelerated by IID. In such cases it can save you 50% or more. I haven't tested it in quite a while, but the last time I did, it was an overall 10% improvement.
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.