Author: Robert Hyatt
Date: 07:21:27 04/22/01
Go up one level in this thread
On April 22, 2001 at 07:48:58, Frank Schneider wrote: >Hi, > >Gromit uses the standard PVS and sometimes after (e.g.) a fail-high >it only gets very short PVs. The reason is that after searching > >failhighscore = -pvs(-beta, -alpha); // returns failhighscore >= beta > >the research is done with > >realscore = -pvs(-infinite, -failhighscore); > >This search may return realscore == failhighscore and in this case >the PV is not saved, because the score is not inside the searchwindow. > >Of course it would be possible to do the research with a window of >(-infinite, -(failhighscore-1)), but this needs slightly more nodes. > >How do others ensure to get "good" PVs? > >Frank There isn't much you can do... this is a problem that first appears when you add hashing. Null-move tends to make worse, but I had this in Cray Blitz even before null-move was used. You can reduce it by clearing the hash table before every new search (not each new iteration) but this hurts efficiency. And it still won't eliminate every case. Best antidote is to use internal iterative deepening so that when you do start the next iteration with no PV/hash information, you still get decent move ordering...
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.