Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Quiescence Explosion

Author: Robert Hyatt

Date: 08:45:02 11/26/02

Go up one level in this thread


On November 26, 2002 at 11:38:40, David Rasmussen wrote:

>On November 26, 2002 at 10:28:24, Sune Fischer wrote:
>
>>
>>You must have some extension that goes mad, like a threat extension or
>>something. I get around 55% qnodes, so you are right it is high but not _that_
>>high.
>>
>
>An extension? They're Q-nodes, and I don't "extend" in Q-search.


Let's get on a common naming convention first.

node = interior node as is searched by Search().

leaf node = _first_ call to q-search (Quiesce()) because those nodes are
not "optional".

quiescence nodes are any _other_ q-search nodes below leaf nodes.

Most count leaves and q-search together, which is fine.  And you should see
numbers in the 50% range there because for every move at the last full-width
ply you search, you will get one leaf node for sure, plus (optionally) more
capture nodes below that...

If you count leaf nodes as q-nodes, you can't possibly get down to 5-10%
and you can see this by trying a tree with a branching factor of (say) 2.0.
3 plies means you try two moves at ply1, 4 at ply2 and 8 at ply3.  For each
ply3 node, you get one leaf node which is 8 leaf nodes vs 14 interior nodes.
Then you tack on captures and you pass 50% instantly...




>
>>>I'm counting like this:
>>
>>Why not do it like this:
>>
>>>Score Search(pos,...)
>>>{
>>>   ++nodes;
>>>   ...
>>>   if (depth < ONE_PLY)
>>>       return Quiescence(pos,...);
>>>   ...
>>>}
>>>
>>>and
>>>
>>>Score Quiescence(pos,...)
>>>{
>>  // no counter here...
>>>   ...
>>>   while (anymore interesting moves)
>>>   {
>>>        ...
>>>        MakeMove(pos,move);
>>>        score = -Quiescence(pos,...);
>>>        UnMakeMove(pos,move);
>>>
>>>        ++qNodes;
>>>        ...
>>>   }
>>>}
>>
>>Then you just have total nodes = nodes+qnodes and the percentage will be
>>nodes*100/qnodes. Cheaper to do 1 add outside the search of course.
>>
>
>That's just as good (or better). But I don't think it can be measured at all if
>I do this or that. And the counts should be the same anyway. It's conceptually
>simpler, though, which is why I might change how I do it.
>
>It's still a mystery why I get 97% qNodes in this position.
>
>/David



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.