Author: Robert Hyatt
Date: 08:52:24 06/05/01
Go up one level in this thread
On June 04, 2001 at 23:52:03, J. Wesley Cleveland wrote: >On June 04, 2001 at 23:13:33, Robert Hyatt wrote: > >>On June 04, 2001 at 17:00:10, J. Wesley Cleveland wrote: >> >>>On June 01, 2001 at 13:59:11, J. Wesley Cleveland wrote: >>>I have been thinking about fail high's and fail low's. If you have a fail high, >>>and you have used a significant portion of the time available for this move (> >>>25% ?), you should just make the move without further searching, as you will not >>>have time to resolve the fail high and investigate other moves anyway. >>> >>>Fail low's are a different problem. When you get a fail low, what you know is >>>that the current move has an upper bound at ply n of the ply n-1 value - window >>>while all other moves have an upper bound at ply n-1 of the ply n-1 value. It >>>seems it might be better to search some of the other moves before re-searching >>>this one. >>> >>>This leads to a crafty question. In search.c, if all the values are < alpha (a >>>fail low), the value returned is alpha and not MAX(values). Why is this ? >> >> >>Why does it matter? The value you get back is totally useless. Because >>everything cut off low on the alpha value. > >Is not the value you get back an upper bound for that subtree ? If you return >and save that value in the hash table, if you later get a fail low closer to the >root, you may not need to re-search this subtree. I don't save root-move scores in the hash table. I can't. Because that would mean before I search the first move at ply=1, I get a hash hit and I don't want that. I want a PV with at least the first move. A hash entry with an UPPER flag has no best move so that would be useless at the root. That value will be saved from the ply=2 search of course... and it will be hit again of needed... but only after a ply-1 move has been searched.
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.