Author: Robert Hyatt
Date: 02:53:36 08/04/98
Go up one level in this thread
On August 03, 1998 at 14:33:17, Roberto Waldteufel wrote: > >Hi Bob, > >Maybe I am misunderstanding what happened, but something sounds a little strange >here. When the search first finds that the pawn can't be held and fails low at >the root, it gives you an upper bound on the position that is less than the >value you previously used for beta by about the value of a pawn, so surely at >this point you re-adjust the search window at the root to take account of this, >so that when you re-search to depths 1,2 etc you do so with lower expectancy. >The positions you reach in the researches should mostly have bounds stored from >the search that failed low to start with, so I don't see why the search should >keep failing repeatedly unless you hit hash records from lower depths (before >the initial fail-low) and allow the lower depth re-searches to fail high and >override your upper bound from the orriginal fail high, but could you not easily >prevent this from happening by disabling the researching process until you had >reached your previous depth again? I would be surprised if Schaeffer is >recommending something as useless as you make it sound. > >Best wishes, >Roberto doesn't help, although I do this. Because if you limit beta, and search from iteration 1 again, but with alpha/beta centered on the expected window at the fail low depth, the first move that fails high at ply=1, and one will fail high, relaxes this. The problem is that you filter new moves back up to depth=8, where the promptly fail low. But it takes time to filter them back up. If you don't let the fail-high happen on the re-searches at shallower depths, you may as well just pick a move randomly to search, bcause *all* will fail high since it seems that the pawn can be held until you reach the right depth. Note that this was the only thing I looked at. There were other positions where it was also way slower... but this caught my attention..
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.