terry mcintyre wrote:
It is possible to get some remarkably high correlation
between the moves played by pros and a predictor - yet
still not have a good program. Why?
We have a random variable, the place at which a player
plays, and some variables that we can compute. The
distribution of the random variable, depends on this
variables. Our variables are shape urgency and life.
Of course, I agree that it would be great to include
life in our statistical model. It would be a much better
model. The problem is we can compute shape in few clockcycles
but we cannot compute life/death. (The revised Benson methods
are of no use in real games because they detect life/death
when it way too late.)
All the pattern logic is not intended to determine what
move has to be played, but only to select candidates for
later methods. Also, it is not about what a player did
once, but what is statistically sound in over 10 M positions.
We ignored life/death when learning and we ignore it
when we just suggest the best moves in terms of shape.
That is fair, but, of course, the interaction between
shape an life would change things for better.
UCT methods are (among other reasons) strong in 9x9
because a random move (not in the first 2 lines in the
beginning) is a reasonable move, and then the stochastic
parts of the search finds good candidates (RAVE and similar).
But in real go, there are 361 legal moves for move 1!
And CrazyStone (as far as I know still the strongest 19x19
program) is "still" 2 kyu. That is the state of computer go
in 2008, not the remark made by a pro about how hard it was
to beat mogo by 9 stones.
Most people in this list seem to be against human learned
shape, but the truth is that we don't know if it is useful
or not. That depends on the price we pay for urgency. In
terms of RAM it is a couple of tenths Mb, that's nothing
today, it was a lot years ago. Some people assume that you
can predict what airplane society would be without building
an airplane. Just move people in a bus, the airplane is the
same only faster. That's not how things work. Unfortunately,
you have to build an airplane to see what it can or cannot do.
Some call it overengineering, but if it is not fast, then
I agree it will be useless almost for sure.
Jacques.
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
