Re: computer-go: Which Maths Is OT?

[Martin Mueller <mueller@xxxxxxxxx>]

David Fotland said:

>I've said this before, but here it is again :)  I don't think there is
>a math to computer go.  I've seen lots of weak programs based on
>some mathematical concept.  Go is an AI problem, not a math
>problem.  So it is a big messy problem, and analyzing it rather
>than coding it will lead nowhere.  Instead of thinking about
>mathematical functions for go, think about knowledge representation,
>automatic learning, move selection for fast reading, move selection
>on the full board, etc.  Think a lot about to quickly and accurately
>evaluate the status or strength of a group, and how to decide when
>you have to read to get a good answer.  Think about lookahead
>algorithms for fast life and death or tactics reading.  Think about
>how evaluate conenctions or territory or eyes.
>

I agree that computer Go is foremost an AI problem, but I think the math is
very handy in some places. Besides being messy at times, Go also has a lot
of geometrical and logical structure that you can exploit. However, I do
not mean some magic influence function or such, but discrete math such as
combinatorial game theory or Benson's work on absolutely alive groups.
These mathematical structures and analyses can serve as a sound basis for
knowledge representation, move selection etc. Much of my own work is in
this grey area between math and AI, so I must contradict strongly if you
claim that all math is useless :)

>
>I think the most interesting topic is machine learning for go.  The
>strong programs all have hand-tuned pattern databases with thousands
>of patterns.  How can you make such a database automatically?
>
>Perhaps Michael Reiss can share his method of buiding patterns for
>move suggestion from professional games?
>

I doubt Michael will share his secrets with you :)
There are some nice Japanese pattern learning programs. Look at the work of
Takuya Kojima and Atsushi Yoshikawa
http://www.brl.ntt.co.jp/people/kojima/research/index.html, and at the two
papers available on
http://www.etl.go.jp/etl/divisions/~7236/Events/workshop98/, "Move
Evaluation Tree System" by Hiroto Yoshii and "Memory-Based Approach in
Go-program Katsunari" by Shinichi Sei and Toshiaki Kawashima. I think in
the meanwhile they have further improved these programs.

	Martin