[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] A chess programmer at the Go-Olympiad

> In 9x9 the 3  Ply searching NeuroGo was clearly better than GA.

That also supports my own belief that ANY search will be a big help.
Even a 3 ply search  sorts out problems.  

The evaluation of a go program needs to be as good as it can be, up to
the point of  diminishing returns.  The you add  global search layers,
even if  it is only 1 or  2 extra.

That might mean that  only a 2 or 3 ply search  works.  In 20 years it
might be a  5 or 6 ply  search.  You don't add depth  for anything you
can fix with evaluation cheaper.

It may be that only 1 ply search works at 19x19 GO right now.  But in
20 years that might be 4 or 5 ply or more.


>    I would offer you a different bet:

>    1. That you will pretty soon give up on the idea of making an
>    evaluation function as fast as you are thinking of now. Really the only
>    thing you can do without doing some tactical reading is s.th. like
>    Benson's algorithm of unconditional life, and this is just not
>    sufficient on 9x9.

>    2. That you won't beat GNU Go on 19x19 within 2 years.

>    3. That after 1 year, I can beat your program on 9x9 giving it a 2 stone
>    handicap on 9x9 after 4-5 training games for me. (I am 4d, and 2 stones
>    is about the handicap I would give to an 8k.)

GNUGO has  come very  far and is  a very  nice program.  But  still, I
would favor  Chrilly at  point 2.  Chrilly  is an  extremely competent
engineer, so  he will sooner  or later figure  out what works.   If he
gets into 19x19 he will write a strong program (relatively speaking.)

- Don







   X-Authentication-Warning: linux.local: arend owned process doing -bs
   Date: Sun, 7 Dec 2003 14:54:03 +0100 (CET)
   From: Arend Bayer <arend.bayer@xxxxxxxxxxxxxxxxx>
   X-X-Sender: arend@xxxxxxxxxxxxxxxxx
   Content-Type: TEXT/PLAIN; charset=US-ASCII
   X-BeenThere: computer-go@xxxxxxxxxxxxxxxxx
   X-Mailman-Version: 2.1.2
   Precedence: list
   Reply-To: computer-go <computer-go@xxxxxxxxxxxxxxxxx>
   List-Id: computer-go  <computer-go.computer-go.org>
   List-Unsubscribe: <http://computer-go.org/mailman/listinfo/computer-go>,
	   <mailto:computer-go-request@xxxxxxxxxxxxxxxxx?subject=unsubscribe>
   List-Archive: <http://computer-go.org/pipermail/computer-go>
   List-Post: <mailto:computer-go@xxxxxxxxxxxxxxxxx>
   List-Help: <mailto:computer-go-request@xxxxxxxxxxxxxxxxx?subject=help>
   List-Subscribe: <http://computer-go.org/mailman/listinfo/computer-go>,
	   <mailto:computer-go-request@xxxxxxxxxxxxxxxxx?subject=subscribe>
   Sender: computer-go-bounces@xxxxxxxxxxxxxxxxx


   Dear Chrilly Donninger,

   > Like every chess programmer I am a believer in search and I do not really understand why Go programms do not use it (Martin gave me some good reasons, but the believe is too deep that one can be converted while drinking one beer). I was therefore deeply impressed by Aya. It seems to be one of the first real searchers. GA had not much trouble with Aya on 19x19, but it was helpless against the 7 Ply search in 9x9. The same holds with NeuroGO. On 19x19 it was a clear-cut win of GA. In 19x19 NeuroGo searches 1 Ply. In 9x9 the 3 Ply searching NeuroGo was clearly better than GA.

   I am also a believer in search. But there are a lot of other things that
   I believe one has to do for a go program as well :)

   > The node count of Go-programs is unbelievable slow. I also thought that the number 620 on Ayas screen means 620 Kilo-Nodes. But Hiroshi explained me, its nodes. My impression from looking at the Gnu-Go and the GA code is. It is not only the complexity of the game. Go programmers obviously do not spend months to save a few nano-seconds in their time critical parts (maybe because there is no time critical part).

   Of course I cannot tell about GA, in GNU Go, we do care about optimizing
   performance critical parts (which is a pretty small amount of the actual
   code) of GNU Go. We don't care much about speedups of about 1% or so,
   but everything higher than that is us worth some effort.
   (But we are not yet doing global search.)

   I would offer you a different bet:
   1. That you will pretty soon give up on the idea of making an
   evaluation function as fast as you are thinking of now. Really the only
   thing you can do without doing some tactical reading is s.th. like
   Benson's algorithm of unconditional life, and this is just not
   sufficient on 9x9.
   2. That you won't beat GNU Go on 19x19 within 2 years.
   3. That after 1 year, I can beat your program on 9x9 giving it a 2 stone
   handicap on 9x9 after 4-5 training games for me. (I am 4d, and 2 stones
   is about the handicap I would give to an 8k.)

   Anyway, good luck!
   Arend

   P.S.: In case someone hasn't heard about Brutus: It has been a remarkably
   successful project as work on it started AFAIK only 3 years ago, and it's now
   competing with the absolute top programs.

   -- 
   Arend Bayer, Flodelingsweg 27a, 53121 Bonn, 0228-9813803
   arend.bayer@xxxxxxxxxxxxxxxxx

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

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