Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Chess Tiger & Fritz - question re nps vs. chess knowledge

Author: martin fierz

Date: 02:56:46 06/21/01

Go up one level in this thread


On June 20, 2001 at 14:25:03, John Hatcher wrote:

>Maybe the answer to my question is obvious, but I'd be interested in what
>programmers have to say on the subject.
>
>I've recently been running matches between Fritz6 and Chess Tiger.  On my old
>Celeron 333mhz computer Fritz6 runs at about 250 knps, and Chess Tiger runs at
>about 80-100 knps.  To my delight, and surprise, Chess Tiger is doing very well
>against Fritz - very well.
>
>Since, as I understand it, "nps" is related (at least nominally) to search depth
>in a given time period, it would seem that, all things being equal, the greater
>the "nps" the stronger the program.
>
>Of course, things are never really equal between two programs.  As evidence of
>that, Chess Tiger with its lower nps holds its own quite nicely against Fritz.
>For this to happen it would seem that Chess Tiger must have some chess
>"knowledge" programmed into it that allows it go toe-to-toe with a program that
>searches 2.5x as many positions per second.
>
>Is the programming of this chess knowledge, even among the programmers of top
>commercial programs, still as much art as science so that the programmer of
>Chess Tiger has found a "smarter", more efficient way to implement chess
>knowledge into Chess Tiger.  I guess I supposed that much of this knowledge
>would be common currency by now.
>
>Sorry for the long-winded question.  I wanted to be clear.
>
>JOHN

in my checkers program, i have various move ordering schemes. if i turn them on,
i get much less nodes/seconds, but i can search deeper all the same. the
same goes for ETC - using it lowers nodes/sec, but has a net positive effect.
i can choose to run my program at 1.1MN/sec without these things or at 750kN/sec
with them, and it's better in the 'slow' mode.
there may also be a big difference in the eval. i know fritz does something fast
but stupid: before starting the search, it assigns pieces values depending on
which square they are. these values remain unchanged during the search (AFAIK).
the result is a fast search but a dumb evaluation - i'm sure they tested it and
found it to be ok, but it can be very irritating at times. it is of course
well-known that if you have a king which is not too safe, you should try to
exchange queens. fritz eval understands this, and gives the attacking side a
bonus - BUT: in the lookahead, it keeps this in, even if the queens are traded.
it will evaluate a line with a queen trade as +1 for the side with the safe
king, but once the queen is gone, it will suddenly say 'equal' - which is
correct, but it didnt see it in the lookahead. i find this very disconcerting
when analyzing games with fritz...

cheers
  martin



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.