Author: Vasik Rajlich
Date: 03:38:42 07/16/04
Go up one level in this thread
On July 15, 2004 at 16:59:14, Dieter Buerssner wrote:
>On July 15, 2004 at 16:43:39, Andrew Dados wrote:
>
>>This above has some implication in replacing schema for transposition table:
>>always replace FL nodes with FH nodes for same remaining depth ('draft') because
>>FL nodes are less costly to compute (cost of FH at depth=d is close to cost of
>>FL at d+1).
>
>Interesting idea. But I am not sure, if it is correct. Assume remaining depth 2
>and average of 40 moves, no extensions, pruning, qsearch. In FH node, you search
>one move, its 40 children, which then will call one eval. In FL node, you search
>40 moves, each time one child, which will call one eval. Not?
>
>BTW. I tried in the past exactly such a scheme. My idea was, that FH scores are
>more useful, because you have a rather reliable move, that will help move
>ordering. FL nodes don't have this (and really don't need moveordering). So FH
>should be a bit more valuable. In practice, it did not seem to make a
>difference.
>
>Regards,
>Dieter
According to my tests, fail-high nodes outnumber fail-low nodes by a ratio of
around 18:1, while the cost of resolving a fail-low node is greater than the
cost of resolving a fail-high node by a ratio of around 15:1.
This suggests that for transposition purposes, it is better to keep fail-low
nodes, than fail-high nodes, of the same depth. In endgames, this will be
especially true.
Regarding move ordering issues (ie quality of fail-high moves vs fail-low
moves), this is a huge topic. Currently Rybka prefers fail-high moves, but I see
no essential reason why this should be the case.
By prefer, I mean that if at the same depth Rybka both fails high and fails low
(ie along the pv), she will store the fail-high move rather than the fail-low
move in the HT.
Vas
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.