Re: [gnugo-devel] Re: [computer-go] SlugGo v.s. Many Faces

[Don Dailey <drd@xxxxxxx>]

I  think this  is  a reasonable  approach,  but it's  good to  reserve
judgement.  As has been mentioned,  the good results against gungo has
a "reasonable"  explanation (even  if it still  remains to be  seen if
this explanation is correct.)

But the  results against Many Faces  is still based on  a pretty small
sample.  About 20  games.  This of course does  not including whatever
has been  played since then and  not yet reported.  Anyone  who does a
lot  of automated  testing knows  that there  is very  little  you can
assume  based  on  such a  small  handful  of  games, even  if  fairly
lopsided.  About all we can say is that it looks pretty good so far.

I would love to see these  results continue strong.  At the very least
it hints  at a way to  move forward in a  way that enables  us to take
advantage of the  massive increase in computing power  that is sure to
come.

- Don



   From: David G Doshay <ddoshay@xxxxxxxxxxxxxxxxx>
   Date: Fri, 27 Aug 2004 10:23:04 -0700
   Sender: computer-go-bounces@xxxxxxxxxxxxxxxxx

   On Aug 26, 2004, at 2:47 AM, wagjan@xxxxxxxxxxxxxxxxx wrote:

   > I am afraid, that there are many situation where the move which
   > seems to be the second best at first sight is much better then the
   > first. (Or the first leeds to a trouble, which is the same... :))

   This is exactly why we consider many of GNU Go's highest ranked
   moves as candidates and then do lookahead on each.

   > And
   > when this happend at 3rd ply, then searching the wrong line in 16
   > deep is worthless.

   This is the "common knowledge" we all learn from our first lessons
   in tesuji and life/death problems. This is undoubtedly true when the
   goal is optimal play for the correct reasons ... a great goal. But I
   think that too many people fixate on this reasoning and neglect to
   note that while the long term goal is optimal play for the correct
   reasons, there is also the short term goal of simply playing better.
   It is clear that this non-optimal once-branched and then linear
   deep lookahead (that we implemented just because it is an easy
   way to start) in fact plays better, even if not optimally, for the right
   reasons, or even with complete or correct information.

   I am fully confident that when we start branching in the lookahead
   paths we will more accurately evaluate the paths and thus play
   better. But once we found that this method worked as well as it
   does, we decided to gather statistics before moving on to other
   methods we assume are better.

   > It is not supprising that it could be stronger by 3-
   > 5 kyu against himselt,

   Because we win about 70% with 5 stones in "self play" against
   GNU Go, I'd say we picked up about 6 stones of strength, but as
   previously noted this is because we correctly predict GNU Go's
   responses. To lessen this effect we have played against GNU
   Go set at level 15 while we use level 10.

   > but it is interesting for me that it is
   > significantly stronger against Many Faces.

   Here we pick up perhaps 3 stones ... the data is still being collected.

   > I guess that 16 ply 'very
   > narrow' search can not help more then 1 kyu against human.

   Time will tell.


   Cheers,
   David


   _______________________________________________
   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/