Author: Gian-Carlo Pascutto
Date: 06:33:44 12/27/01
Go up one level in this thread
On December 27, 2001 at 09:26:46, Uri Blass wrote: >I understand from other posts that it was fail high that >was not confirmed but I still do not understand how fail high >can be wrong. > >I read in bruce pages that it happens only to strong programs >and I also remember from previous discussion about MTD that >it is related with hash tables. > >Is the reason for the fail high that were not confirmed only >null moves or also hash tables? Both individually can cause it. With nullmove, the reduced depth search can hide a threat, and cause the search to fail high when it should not. When researching with a different window, the nullmove score can be too low to get an immediate cutoff from the reduced depth search, and the full search can fail low instead because of a threat that the reduced depth search had not seen. With hashtables, it is possible that a deep entry was left over in the hashtable because of pondering, and that it had already gotten overwritten before the research reached the relevant position. The result from a shallow search has to be used, which can have a (very) different value than the deep result that caused the initial fail high. >I am also interested to see the tree that cause >to this problem I cannot trivially construct one, sorry. >chessbase interface can understand tree so if it is possible >to post a practical tree as small as possible >that cause the problem with comments after every node >(like number of node,alpha value,beta value) >then it may be productive for me to look at it > >I think that a program that translate nodes of a program to a tree that >chessbase interface can understand may be productive >for chess programming because I believe that >looking at the tree node after node >may be productive to get ideas. Remi Coulom has an excellent utility on his homepage to browse through large search trees. I have not been able to get my program to output data it can handle, though. -- GCP
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.