Re: computer-go: Fast scoring program?

[Erik van der Werf <E.vanderWerf@xxxxxxxxxxxxx>]

> > also try limiting yourself to 5x5 or 7x7 boards to accelerate the
> > learning process.
> >
> >
> 
> Ahh but if it gets good i want to put it on the ladder so 9x9 is size i
> need
> 

Limiting yourself to small boards not only accelerates the learning
process. It also allows you to evaluate the learning algorithm much
better. For example, your networks should in principle be able to always
win as first player on 3x3 boards. Ofcourse this may seem trivial, but
you might be suprised how easy a trained feedforward network can be
fooled by unorthodox play, even on such a small board size.

The basic problem you are facing is the lack of knowledge of the
topological structure of the go board. The feedforward nets will have to
learn all spatial relations from examples, which might be extremely hard
on 9x9 boards.

Good luck,
Erik