Re: computer-go: Programs learning to play Go

[Zachary Tellman <mordochbob@xxxxxxxx>]

    I suscribed to this list on a whim more than anything else, however
this thread has caught my attention.  My knowledge of computational go is
limited, however I do know something about neural nets.  Much of the
discussion seems to center around training a net to do the same thing a
scoring algorithm would in any other go program.  This seems to be a
somewhat limited application for a neural net, as firstly neural nets can
only approximate their I/O sets (they will never actually acheive perfect
accuracy), and the means by which one can score the game have already
been established.
    I fail to see why someone hasn't yet attempted to use a neural net
to, say, determine the opponents response to a move or something on that
level of abstraction.  The problem as I see it is that the only
development done so far (or at least the development discussed) has been
using the basic backpropagation algorithm.  Sure, you can *try* to train
a network using thousands of I/O sets which individually consist of
hundreds of data units.  But as was said before, this has a snowball's
chance in hell of working.  But that doesn't mean that simplification of
the data is the answer, because that allows for perhaps a fairly major
misrepresentation of the game.  Instead, perhaps some looking at
alternative neural net algorithms is in order, namely SOMs and variants
thereof.  This variety doesn't necessarily spit out a number
automatically (it usually requires an additional algorithm of some sort
to analyze it), but it may well be a valid alternative.
    This topic most certainly should be explored in more depth, and I'm
curious what everyone else has to say.

Zachary Tellman