Author: Bruce Moreland
Date: 10:14:43 04/01/98
Go up one level in this thread
On April 01, 1998 at 03:21:58, Andreas De Troy wrote:
>[I wrote a similar post a few days ago in rec.games.chess.computer, but
>it did not show up there (at least I did not see it), so I repost it
>now.]
>
>I am working on My First ChessProgram and I wonder how you people count
>these nps (nodes per second). I see that Nimzo98 claims 200 000 nps on a
>Pentium 200 and when I compare this to *my* current program I cannot
>imagine how I could ever reach, say, half of this number. I got
>typically between 45 000 and 70 000 nps with material evaluation only.
First remember that NPS, while it is an easy number to measure, doesn't
necessarily have an exact correspondance with Elo. It makes little
sense to optimize for it in absence of other factors. For instance, if
you can order your moves better, so that you search 20% fewer nodes to
get to depth D, but each node takes 10% longer to search, you'd be a
fool not to do this even though everyone would say you've gotten slower.
But on the other hand, when people start out, they all go slow by any
metric.
Nimzo probably does very little evaluation at the tips. You might want
to compare yourself to something more normal, like perhaps Crafty.
If you stay interested in this subject, you will probably end up
rewriting every time-critical function in your program a few times, this
will make you go faster.
>So how do you count this? I saw in Crafty the "nodes++"-statement at the
>beginning of the ab-search and the q-search, and that's how I count too.
>But maybe you should count the number of generated moves instead? (The
>number would be much higher then.)
If your car has a speedometer that measures miles per hour, it doesn't
go faster if you put a new one in that measures kilometers per hour.
>And a few days ago I saw in a message from KK that hashing influenced
>this number too so my second question is: are the hits counted, or
>aren't they?
In my opinion, they are counted. One natural way to write a search
would be as follows:
search( ... )
{
nodes++;
if ( ... ttable match ... )
return value_from_ttable;
while (more_candidates) {
...
search();
...
}
return value_from_search;
}
It might be possible to moe the transposition table probe, so that it is
not at the top of search, but rather it is executed right before you
recurse into search:
search( ... )
{
nodes++;
while (more_candidates) {
...
if ( ... no ttable match ... ) {
search();
...
} else
nodes++;
...
}
return value_from_search;
}
In order to get the same node counts for the exact same work, I inserted
a new "nodes++" statement in the above, and I think it would be "fair"
to do this.
bruce
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.