Author: Simon Finn
Date: 06:15:16 09/16/02
Go up one level in this thread
On September 14, 2002 at 23:41:55, scott farrell wrote: >Does anyone know what implications fail-soft might have on hashing? > >If you do a search say from 1 to 2, and under fail soft it returns 2.5 >(fail-high / beta cutoff). I am currently storing the score as a beta cuttoff of >2.5. If I see this position again, and the depth is good enough, I return the >value from the hash table, so this is fail-soft as well. > >if (mainTable.flag[te] == flagBeta) > if (mainTable.value[te] >= beta) > return mainTable.value[te]; > >I was tracking IID failures, and came up with some debugging that shows >occaisonally no entry in the hashtable after a score of >= beta. Is this surprising? If you arrive at the same position by different routes, you may have different alpha-beta windows. This means that a fail-soft value (generated with a narrow window) is not necessarily reusable (for a search with a wider window). When you search a node at greater depth, you may also have a different alpha-beta window (because preceding nodes, also searched at greater depth, now return different values). When you start IID from a node, you may encounter a high-draft/small-window/fail-soft node (generated by IID from some other node) in the hash table. Depending on the actual fail-soft value, this may or may not lie outside the current alpha-beta window. If the stored fail-soft value is within the current (wider) window, you will get a "hash table failure". Does this explain what you are seeing, or have I misunderstood the problem? Simon >I am still >trying to trace it down. It seems lots more of these occured nearer to the >terminal nodes (search horizon). I stopped IID recursing nearer the terminal >nodes, and many of the failures stopped. I guess I might not be hashing enough >details near the terminal nodes. I have tried turning on and off storing values >from the qsearch, but it didnt seem to effect things.
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.