Re: [computer-go] Pattern matching - example play

[William Harold Newman <william.newman@xxxxxxxxxxx>]

On Sat, Dec 04, 2004 at 12:29:34PM +0100, Heikki Levanto wrote:
> On Sat, Dec 04, 2004 at 04:24:10AM +0100, Vincent Diepeveen wrote:
> > 
> > It is not a game tree search like go or chess are.
> 
> No, but among all the differences, there may well be a few similarities.
> 
> Backgammon has a huge branching factor that makes search difficult.
> Evaluating backgammon positions is difficult (because of the statistical
> nature of the game).
> 
> Go has a huge branching factor, and a difficult evaluation.

I think the evaluation difficulty may be different in Go. It is a
common property of many approximation and learning methods (not just
of NNs) that smooth functions tend to be easier to approximate than
jagged functions. It doesn't seem to be possible (considering the way
that people have shown how to embed NP and harder problems in the
score of Go) to define a useful metric where nearby Go positions tend
strongly to have similar scores. As I understand it (remembering the
rules only vaguely, so correct me if I'm wrong...) the situation is
much different in backgammon: there, it is difficult to find a
position where winningness is changed dramatically by a small move of
one piece.

-- 
William Harold Newman <william.newman@xxxxxxxxxxxxxxxxx>
In an election or more general political debate, there's a further
point to be made: when assessing people's beliefs, you have to listen
to what they say, not just play actuary and assume you have them
figured out. -- http://www.iwamatodjishi.com/archives/000242.html
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/