Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: NPS

Author: Robert Hyatt

Date: 05:16:34 11/14/97

Go up one level in this thread


On November 14, 1997 at 04:15:24, Chris Whittington wrote:

>
>On November 14, 1997 at 03:30:27, Andreas Mader wrote:
>
>>Chris has claimed many times that his CSTal only searches 4000 NPS and
>>performed very well in Paris, whereas the 'fast searchers' (>100
>>KiloNPS) had their problems.
>>
>>This is of course true, but it leads to a question: How do you count the
>>NPS? I don't know if this has been asked before, but I would like to
>>know, how different programs are counting NPS. Can it be that Chris is
>>simply the 'king of understatement' because he uses a different
>>algorithm?
>
>I'm pretty sure I count in the standard way; here my method, Bob or
>whoever can say if its standard or not .....
>
>CSTal tree searches, like everybody else.
>
>At each node it does the following:
>
>
>makemove
>evaluate
>if (extend_this_node)
>{
>  do a hash lookup - it might find a mtach and cutoff here
>  go one more deeper, recursively and return back to here
>}
>unmakemove
>if score>alpha etc
>
>the above process count for ONE node.
>if the hash finds a cutoff it still counts for ONE node
>if the node gets extended then we add more nodes by the same process
>

basically my search/quiesce functions look like this:

int search() {
  nodes_searched++;
  ...
}

so I think we are counting the same thing.



>
>>If you find a position in the hash tables, do you count it?
>
>yes, as one node, as described above

ditto here, because the nodes_searched is incremented upon entry to
search, *then* I do the hash probe and a successful one will return
with no further searching.  I do *not* count the nodes that the hash
hit replaces, because I don't store that info in the table (although
I suppose I could.)



>
>>If a position has a 'clear' material balance so that you do not need to
>>do some 'fine tuning' in the evaluation function, do you count it?
>
>CSTal doesn't do a lazy eval, since the eval function can produce very
>large scoring swings. Only MATERIALISTS do lazy evals :)

you *could* do one.  IE how big a positional score can you produce?  I
can produce one of up to at *least* +/- 10 pawns, and in simple endings
+/- 20 pawns...  It only means as the circumstances change, I begin to
be less "lazy".  This is why, on a P6/200, I search 60K in the opening,
80-100K in the middlegame, and 120K in the endgame.  Because I have to
get less lazy as the game progresses because positional scores can get
bigger and bigger...


>
>So not applicaple.
>
>>When you do quiescence search, do you count the NPSs?
>
>CSTal doesn't work to this MATERIALISTIC algorithm. Any node, extension
>or whatever you want to call it gets counted.
>
>>etc. etc.
>>
>>Could it be that the 'real' factor between knowledge based programs and
>>fast searchers is not that big (4000 : 200.000)?
>
>Don;t think so, but comments welcome ....
>
>Chris
>
>>
>>Best wishes
>>Andreas



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.