Author: Ernst A. Heinz
Date: 09:39:04 04/07/00
Go up one level in this thread
>>>No, razoring is when you do futility pruning at the last ply of your full-width >>>search. I think most people don't like it. I tried it and all my test suite >>>scores went way down, even when the futility threshold was pretty conservative. >>>It just causes to the search to miss too much... >> >>I thought that was called "extended futility pruning" and razoring was >>simply reducing the depth of the search by 1 ply if nothing interesting >>was going on near the leaf and the move made does not bring the score >>near alpha. > >You are correct. "Extended futility pruning" is what Tom described, i.e. if >the following condition is met at pre-frontier nodes (depth 2) then prune; > >mat_balance(node) + mat_gain(node)+ extd_futil_margin(node) <= alpha >(from Heinz, 1998) > >In general, the efm() is about a rook to a rook+pawn. > >Razoring similar to what you described, but an enormous material unbalance is >generally required, i.e. a queen's value or more. The search depth is reduced >by one ply. This is normally applied at pre-pre-frontier nodes, i.e. depth 3. > >I do agree with Tom, though, that in my experience, both these methods, and >also normal futility pruning, vastly reduce tactical accuracy. I know that >Ernst will disagree, but that's just my experience. Yes, I strongly disagree because normal futility pruning is _theoretically sound_ for searchers with a capture-only quiescence and a futility margin that is at least as large as the maximal positional score produced by the static evaluation function. Standard futility pruning simply "lifts" stand-pat cutoffs from horizon nodes to frontier nodes. In your description of extended futility pruning it should read "mat_gain(move)" and not "mat_gain(node)", i.e., the scheme only cuts off outgoing *quiet* moves at potentially futile positions. This is very important -- the search still follows all others including checks! If you do not restrict the pruning to quiet moves, you end up doing real razoring which is clearly bad and produces exactly the results Tom and you mention. BTW, several strong commercial chess programs use normal and extended futility pruning meanwhile with good success. So, somebody must do something wrong ... =Ernst=
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.