Learning and Miai

[Wolfgang Werner <wolfgang@xxxxxxxxxxxxxxxxxxxxxxxxxxxx>]

Hi there,

(1) Learning
It seems to me, that in writing a good GO-program a lot of work is done in
trying to transfer the knowledge of an experienced GO-player into the
machine. Due to the immense tree width and the (practical) impossibility of 
evaluating all the positions correctly and timely, I think a learning 
approach is the only solution for a real good program. But, for every 
learning approach there must be a kind of a-priori information on what and
how to learn. In my opinion, the best approach would be to minimize this 
a-priori information to get a most flexible learn mechanism. 

Has anybody already tried such an approach ? And, how much a-priori 
information was needed ?


(2) Miai
Are there any approaches in getting a grip on miai situations ? Every good 
GO-player knows about miai, but how can you tell a program not to play a
promising move, because there is another equal move which brings about the same
effect. So, the program could take sente and play elsewhere (maybe a less
promising) move. The problem with miai additionally is, that the situation 
need not be a local life or deatch problem, but e.g. connect or live, or
invade or reduce.

Any comments ?


Wolfgang

----------------------------------------------------------------------------
-------
Dipl.-Inform. Wolfgang Werner		 Phone +49-89-6004-2545
Institut fuer Erdmessung und Navigation	 Fax   +49-89-6004-3019
University FAF Muenchen			 mailto:wolfgang@xxxxxxxxxxxxxxxxx
D-85577 Neubiberg