[computer-go] What time controls should be used for Go?

[Don Dailey <drd@xxxxxxx>]

It seems that ANY time control is arbitrary.  It hardly matters to the
computer.  When we program a computer to play a game we tend to use
time controls that are convenient for us.  Since computers are here to
serve us and not the other way around, I think this should be how we
think of what is natural and what is artificial.  But technically
nothing is particularly "artificial" about any time control.  It's
just the human element that we judge by.

No matter how fast computers get or will ever get, there will always
be something that we can't quite get done without a "little bit more
power."  Most of us learned long ago that the need for "just a little
more computing power" was always desirable, but a bit like chasing
your own tail.  There is, and never will be, as much as we would like.
When we get it, we will find a way to use it and then be disappointed
that we don't have more.

It seems as if the SlugGo team has bumped into this phenomenon and not
recognized it for what it is.  The obvious solution is to wait for
faster hardware, because the wait will be very short.  Meanwhile they
can work on other refinements and improvements.

What they have discovered is scalability.  But the only impressive way
to deomonstrate scalability is to have a system that has been scaled
up!  This means they will probably always wish they had more in order
to compete and the hardware will never be quite adequate for
competition.  When it does becomes adequate, the others will have
tuned up their programs and so SlugGo will feel obliged to reciprocate
with a change in their algorithm such as doing a deeper search
complaining that it plays worse if they don't get this.

But if we let them compete with single processor programs that have
been designed only to play very quickly, then there is little to be
learned that we don't already know and they helped prove, (that go
programs can be scaled.)

Nevertheless, I am quite excited by their discovery.  It is definitely
enlightening to see an existing program such as GnuGo being scaled up
by search.  It's particularly impressive as GnuGo wasn't specificially
designed to be scaled in this way.

I would personally like to see them compete but I would wish for it to
be on an equal footing.  It doesn't bother me the least that they have
a lot more hardware and I don't consider that unfair at all.  I like
it that you can bring any hardware you can muster, you certainly
wouldn't ban an athlete from a competition because he happened to be
fitter than the competition.  But it wouldn't make any sense to give a
sprinter (who may also be one of the fittest athletes) a 5 second head
start in addition.   To prove what?   

I think the big problem with SlugGo is that they are using GnuGo as a
base and it takes GnuGo, a serial program, a certain amount of time to
compute a move and a score.  Again, a short wait for faster hardware
might solve this problem, but by then a newer version of GnuGo might
play significantly stronger but require more time and they won't want
to use the "fast" older version.  They will probably discover that
they can't really compete with the newer upgraded programs unless they
use the new but slower GnuGo.  So they are caught up in the same
vicious cycle we all are, wishing we had more power!

Incidently, my own program is scalable and would benefit enormously
from a long time control tournament.  So I am not being critical just
because I believe that long time controls would hurt my program in any
would should I ever decide to compete.  It would be in my best
interest to argue for the longest time control possible (assuming I
ever compete), knowing that I would essentially be campaigning for an
advantage over the other programs.  But my experience with computer
chess indicates that shorter time controls, from the human standpoint,
seem to be in the best interest of promoting the game and making the
tournament fun and interesting.

Having said all of that, there would be one big advanatage to long
time control tournaments.  It would encourage programmers to think in
terms of scalable algorithms that automatically respond to better
hardware.  It seems crazy to me that a program gets marketed that
might play too slow if I have an old computer, and will play too fast
to take advanatage of the next new computer I may get in a couple of
years.


- Don





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