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

Re: [computer-go] how to use GTP in place of GMP



I understand.   The time controls for  Go are based  on the tournament
tradition of 1 hour per player.   If tournament time was 4 hours would
there be adjustments you could make?


>  Can your program beat Many Faces of Go with a 1000 to 1 time advantage on a
>  19 line board?

No.  At least I  doubt it.  I need an enormous amount  of time just to
get close to gnugo and I believe your program is signficantly better.

How does your program do  against gnugo?  Is it reasonably competitive
or does Gnugo have no chance?   If your program played 1000 games with
the latest Gnugo, do you have a sense of what the score would be?  Has
anyone benched this?

- Don



   From: "David Fotland" <fotland@xxxxxxxxxxxxxxxxx>
   Date: Fri, 13 Aug 2004 19:27:09 -0700
   Content-Type: text/plain;
	   charset="us-ascii"
   X-Priority: 3 (Normal)
   X-MSMail-Priority: Normal
   Importance: Normal
   X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2800.1409

   The commercial programs all struggled to complete a tournament game in the 1
   hour time limit.

   Most of the commercial programs were written back in the days of the Ing
   contest with the
   $1 Million prize.  A program that completed all it's moves in an hour back
   then on a 166 MHz
   Pentium will seem very fast today.  I don't know anyone who is trying to
   make a program
   move in less than a second.  My tactician was designed to answer all the
   interesting tactical
   questions in one hour per game, on a 66 MHz computer.  I don't see any need
   to recode it to
   run slower now :)  The strategic questions can't be solved by any amount of
   search.  I would use
   more time if I knew how to use it to make the program stronger.

   If I use the extra time to do extra searching on moves that the programs
   knows are stretegically
   weak, there is a chance that some anomaly in the evaluation or the search
   will pick a worse move.

   Can your program beat Many Faces of Go with a 1000 to 1 time advantage on a
   19 line board?

   Regards,

   David



   > -----Original Message-----
   > From: computer-go-bounces@xxxxxxxxxxxxxxxxx 
   > [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx] On Behalf Of Don Dailey
   > Sent: Friday, August 13, 2004 7:18 PM
   > To: computer-go@xxxxxxxxxxxxxxxxx
   > Cc: computer-go@xxxxxxxxxxxxxxxxx
   > Subject: Re: [computer-go] how to use GTP in place of GMP
   > 
   > 
   > 
   > Hi David,
   > 
   > I  agree, playing well  should be  the goal.   Computer Go  has always
   > seemed odd to  me because it seems like all the  programs are 
   > fixed up to play really  quickly.  Everyone complains how 
   > abouot  how much they stink, but they  whip out moves 
   > instantly, almost as  if they know all they need to know to 
   > play a move.
   > 
   > There  is a  different mentality  about this  than in  
   > computer chess, probably because  computer chess responds  
   > very naturally to  a highly scalable algorithm (called global search.)
   > 
   > So if  your program  plays just slightly  better than mine,  
   > but yours takes 100 times longer to  compute, is your program 
   > really better?  In computer chess the  answer is no, but in 
   > computer  go the answer seems to be YES.
   > 
   > My  program takes  a  very long  to play  a  good move  (good 
   > being  a relative term of  course :-) It's very scalable, so 
   > I  can lose to any program and beat  any program if we can 
   > simply  ignore the clock.
   > 
   > Go programs  care a lot  about time,  but it seems  to be 
   > only  in the sense that it is a sin to take more than a 
   > second to move.
   > 
   > Of course I'm not familiar with  all programs.  I am curious 
   > about the commercial programs, do they all play "hair on 
   > fire" speed go?
   > 
   > - Don
   > 
   > 
   > 
   >    Programs are already disqualified for being slow: it is 
   > now the norm 
   >    that
   >    a program must play a game in an hour and I cannot get my 
   > code to do
   >    that even though I run on a cluster of very fast CPUs. So 
   > at this time I
   >    cannot participate in any competition. But given the 
   > present paranoid
   >    climate, I cannot compete anyway because I am not even going to
   >    consider packing up my cluster for transport to a 
   > tournament site. The
   >    only solution I can think of is to host a tournament at my site.
   > 
   >    With the present state of Go programming I think that just 
   > being able to
   >    play well should be the goal, not to play well and fast, 
   > but that is 
   >    just me.
   > 
   >    I view the biggest security issue to be sneak attacks on the other 
   >    computer.
   >    For this a referee program in the middle to inspect and buffer the 
   >    communication
   >    stream seems like the best idea. Our group has just 
   > (today) started  on
   >    such a program. If it proves of interest, and I imagine it 
   > might given 
   >    the
   >    volume of email on these subjects the last few days, we 
   > will gladly make
   >    our code available in the normal GNU public license sense.
   > 
   >    Cheers,
   >    David
   > 
   > 
   >    On Aug 13, 2004, at 2:39 PM, Richard Brown wrote:
   > 
   >    > William Harold Newman wrote:
   >    >
   >    >> Another possibility, perhaps only marginally practical now,
   >    >> but probably more practical in ten years, might be to 
   > play games with
   >    >> an average time allowance so short (perhaps 250 
   > milliseconds per move?
   >    >> 150?) that humans can't react fast enough to be helpful.
   >    >
   >    > Hmmm.   Not merely impractical, but highly unfair to my program,
   >    > which has to do some disk access, because the pattern database is
   >    > too large to fit into memory.  That disk access makes my program
   >    > slow, and your proposal punishes me for that.
   >    >
   >    > Is such a program to be disqualified, just because it's "slow"
   >    > in your opinion?
   > 
   >    _______________________________________________
   >    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/
   > 

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