Re: [computer-go] Purpose of Neural-Nets?

["Robin Kramer" <robin@xxxxxxxx>]

> Two of the major motivations for neural-nets in Go are following.
>
> 1. Most of the important concepts in Go (at least as used by humans) are
>  fuzzy by nature. Thus, the fuzzy-logic capability of neural-nets make
> them very  attractive. This situation is due to the fact that no
> effective evaluation  functions exist.

If by fuzzy logic, you mean the ability to generalize, then there are
several algorithms which also have this ability.  I think fuzzy logic is
not essential since the input contains no floating point values.  Floating
point values represent problems for many learning algorithms.

> 2. Knowledge database is very important in a Go program. The capability
> of  neural-nets to extract information make them very interesting.

Neural networks are somewhat less interesting, because the information
that is contained in the neural-net is not exactly accessible.  Compared
to bayesian analysis, decision trees, K nearest neighbor or SVM's, where
the information that was learned could be verified by a domain
expert(professional player).

3. Don't forget that neural networks also have the ability to learn
concepts which are not linearly seperable.

Decision trees, SVM's, and KNN also have this ability.

Also for our "Thought experiment"  is there a preprocessing limit, or a
limit to the amount of human tweaking of the algorithm prior to playing?

Sincerely,

Robin Kramer



_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go