Re: [computer-go] Using floating point sounds very strange to me

[Don Dailey <drd@xxxxxxx>]

I did an experiment just a few years back on the effect of speed on GO
programs, but I don't know if/how the results apply to "real" GO
programs.  I made up a version with a trivial evaluation function and
it searched like a chess program using global alpha/beta.  The
evaluation functions were trivial things like simply counting stones
in one version.  Another version tried to be sophisticated, counting
territory and trying to resolve groups with small numbers of
liberties.

But in each case, I got very chess-like improvements for increasing
the number of nodes searched.  I did most of these runs on 9x9 boards
but did one verify with a 19x19 board.  In short, any evaluation seemed to
scale.

I present this for what it's worth and I don't know what it's worth.
With these evaluations those programs are far weaker that the good
programs are and perhaps the results don't scale, but I have no real
reason to believe that they won't.  The only problem is that those
programs were so weak that no concievable computing system in the near
future could get them up to the strength of already existing programs,
but that doesn't mean one with a much better evaluation wouldn't work.

- Don
 




   X-Original-To: computer-go@xxxxxxxxxxxxxxxxx
   From: "chrilly" <chrilly@xxxxxxxxxxxxxxxxx>
   Date: Tue, 18 Jan 2005 12:25:16 +0100

   >
   >Perhaps the big difference is the maturity of Chess programs v.s that
   >of Go
   >programs. The literature on what has been tried and what has worked
   >likely
   >results in the top chess programs having more in common than the present
   >top Go programs, so a few percent difference in speed shows more
   >distinctly
   >in the chess programs. Right now Go programs are using a very wide
   >variety
   >of algorithms and evaluation functions, and those details seem more
   >important
   >than speed right now.
   >
   Exactly. Chessprogramms are basically all the same. I do not know any
   internal details of Shredder, but I sent recently a testgame Hydra-Shredder
   to Stefan Meyer-Kahlen (Shredder-Programmer) with comments what went wrong
   in this game for Shredder. Stefan confirmed that my diagnosis was perfectly
   right. If everybody does the same, doing it a little bit better is very
   important.

   >
   >Inefficient if what you think you need is all CPUs cranking full speed
   >all the time.
   >
   It is in my understanding a relative simple parallel algorithm. But to make
   it really efficient one would to have change the whole programm. Seems to be
   in Go very difficult. The chess model is much simpler, because there is a
   relative clear-cut seperation of search and evaluation. But it is even in
   chess extremly difficult to keep a network of 32 (or even 72) CPUs busy and
   do avoid unnecessary work at the same time.
   I mention the number 32, because I am working currently on a new version of
   Hydra for a 32-Processor Cluster.
   >
   >I do not follow your line of thought here, particularly the jump back
   >to floating
   >point considerations. I would like to understand what you are thinking
   >...
   Sorry, this was a rather baroque Austrian sentence. If you see pictures of
   Vienna or Salzburg with baroque buildings. It is not only the buildings, our
   way to say things is the same.
   Short version: SlugGo shows, that speed is important in Go.

   One can argue, that this is anyway known. But from my experience at the
   Go-Olympiad 2003 it is known differnently than in chess. E.g. most programms
   did not use the full time-control. The time control is with 1 hour for a
   19x19 game rather fast. The argument seems to be. If the programms cant use
   the time anyway, why sitting unnecessary around waiting for moves. Waiting
   15 minutes for a move like it happens in chess is really a pain in the ass.
   There were also some plans that I build together with Peter Woitke (GoAhead)
   a new Go-Programm. One of the basic requirements of Peter was, that the
   programm strength is monotonically increasing with hardware-speed. I did not
   understand the requierement at all, because this is completly obvious in
   chess. But it is not the case in the current GoAhead.

   Chrilly


   _______________________________________________
   computer-go mailing list
   computer-go@xxxxxxxxxxxxxxxxx
   http://www.computer-go.org/mailman/listinfo/computer-go/

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/