Re: Fuseki and Joseki Database with Neural Network

[Jürgen Kahnert <Kahnert@xxxxxxxxxxxxxxxx>]

Ok, here a summary:

A neural network is unfit to hold a joseki-database, because it is not
exact and much more.

It's nearly impossible to get a corrcet joseki sequence out of a NN.

It's extremly difficult to learn a new sequence without loosing the
knowledge of the old ones.

Nobody knows how a NN learn it's knowledge, so we couldn't decide if it
works for other situations, or in other words, we have to proove
untrained joseki sequences and decide for each if it works correct - and
if not, we have to let it learn all the patterns new (with other
parameters). I think thats not an adequate way of spent my time for a
good joseki database.

The calculation time to train a neural network is immense, but i think
not really an obstacle, because the computers are so powerful and when
we let it train a NN in the background (with a low priority task) no one
would even recognize that there is something in the background that
steels computer-power...
And... who do not switch off the computer when going to work or
something like that? What an immense waste of calculation power (i'm
getting headach only by thinking about this... ;-)

So, what about a hybrid system? A really good an big normal joseki
database combined with a NN to get inputs for choosing where to
concentrate the search, maybe by no suitable move in the database or by
choosing the correct or best alternative of many possible ones.
Maybe a NN would also be able to support the decision where to
concentrate the search in the middle game.

	Juergen