[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: Discarding the rubbish moves
>
> What rules are there for vetoing moves? I'm interested at two levels:
> 1. Proveably bad moves
> 2. Pro level at 99.9%
When learning from existing games, type 2 is easier to learn in some
sense. One only has to take a bunch of pro games (collections of "good"
moves) and a set (of approximately the same size) of random moves (which
are probably bad moves). Using these (position,move)->value examples, a
learning program can learn a theory which eliminates most "bad" moves
without reading ahead. When learning to recognize good moves, one should
take into account that this gives only a first elimination of bad ones.
e.g. moves that are good shape but do not work will probably not be
eliminated.
learning type 1 moves is more difficult as it is not so easy to get good
examples. Moreover, "learning" in the usual sense means that a theory is
built that statistically gives good results most of the time and
"provable" isnt compatible with that.
Still, one can generate examples for type 1 e.g. with life and death
problems for which the correct solution is supposed to be proven.
> As an example of type 1, in a three-in-a-row, the end points are always
> bad moves:
> ?####
> #abc# ("?" and rest of board are "don't care")
> ####?
>
This kind of things can be easily learned by a program.
my program e.g. catches this situation in a more general rule
if [the stone you put will be in atari]
then
if [not a throw-in like situation]
then
the move is probably bad
(the learned tests for throw-in like situations are not so comprehensible
so i dont list them ;-( )
jan