Author: Dave Gomboc
Date: 00:00:12 05/20/99
Go up one level in this thread
On May 19, 1999 at 02:32:31, Ed Schröder wrote: >>Posted by Rémi Coulom on May 18, 1999 at 03:56:14: > >>>Extensions (checks) : 77.979 (13%) >>>Extensions (captures) : 5.807 (1%) >>>Extensions (king safety) : 1.180 (0%) >>>Extensions (on depth) : 7.594 (1%) >>>Extensions (remaining) : 1.094 (0%) >>>Extensions (total) : 93.654 (16%) >>> >> >>Could you explain what "king safety", "on depth" and "remaining" extensions >>consist in ? > >Extensions (King Safety): when one of the kings is in danger an extra >ply is taken. > >Extensions (on depth): on depth means the ply before Q-Search. Various >extensions (extra plies) are tried to avoid a horizon effect. To name a few: >#1 if the last move attacks 2 pieces. >#2 if the king is in check. >#3 if the last move is capture that gains material (static evaluation). >#4 cases like bba2 wrb3 wrf7 >#5 piece attacks like h2-h3 attacking the black knight on g4. >#6 moves that escape from (#5) such as Nf6, Nh6, Ne5 >#7 certain types of mate threats recognized by the evaluation function. > >All of this is safe-guarded with several forward prune techniques to >avoid the tree to explode. Next, extensions on depth is limited to take >2 or 3 of such extra plies also to avoid the tree to explode. It has been suggested to use a limiting factor like "not more than 1 ply extension per ply" or "not more than 2 ply of extension every two plies". Do you find a need for such a limitation after your forward pruning is done? >Extensions (remaining): A number of the list of "Extensions on depth" >is also tried in the tree. Here I use fractional increments of 1/4 ply. One >idea is to push the hash table entries of best lines. In other words, if >the current move is from the hash table and the move of the previous >ply too then the depth is increased with 1/4 ply. Also some promotion >extensions are tried which increment may vary on their importance. "Pushing the hash table" sounds interesting. I have never heard of this being done before! In "Search Control Methods in Deep Blue", a small (I think 1/4 ply) extension was done if a move by the opponent was made possible by the last move of the opponent (i.e. they are following some plan... and might be putting together some kind of combination, even if you're not sure what it is yet). They referred to this as "influence", I think. I do not remember if they invented it or if it was cited from somewhere else. >Ed Schroder Dave
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.