Re: Fuseki and Joseki Database with Neural Network

[Jeffrey Greenberg <jeffg@xxxxxxxxxx>]

Some earlier work I did is on my web pages at:
    www.cris.com/~jgberg
The papers there are pretty casual, but I can add data if people need to be
persuaded more.
The gist of it is that I tried to see how the combination of GP solving power
and GO complexity would play against each other.  I created a very simple
language having only to do with occupancy of points on a board.  I attempted to
evolve individuals that could create patterns of local play based on occupancy
alone.  Each individual was pitted against a database of expert games from which
a set of boards and the expert's play was provided.  The most fit individuals
would match the greatest number of plays.  It turned out that the system was
very ineffectual at evolving a player that could match the fitness test cases,
let alone be tested for robustness.  I did an analysis to understand why this
was.  My conclusion was that even with what appears to be a relatively simple
problem, there is in fact a very large solution space such that (even) GP was
ineffective.  The key thing for GP is to develop a language that encapsulates
enough of the complexity such that the solution space is tractable.  With the
very simple language, it isn't.  The next step then is to enhance the GP
language.

My goal in all of this is to find a way to develop software without requiring
the programmer to be so much of a domain expert.  Using raw GP (in its current
state) and almost no knowledge will not get me there.

Since then I have developed a portable Go Development Environment with enough
expert information to play the game legally.   I am at the crossroads of either:

    1. enhancing & developing my GP engine and integrating into the GP language
more Go knowledge,
    2. choose another technology other than GP alone
On the plus side of the 2nd path is some interesting results from S. Willmott
using goal-based search on Go, ANN work by Mueller and others that has promise,
and there is some interesting technology coming from other areas which I don't
want to point out just yet...
For GP there are some interesting ideas coming from economics that might help
cooperation amongst individuals and help with achieving better convergence (E.
Baum).  Plus the application of strong typing,  demes on parallel hardware,
evolvable hardware, and machine instruction-level GP are all applicable.

There is also some non-GP but EP kind of stuff on Go in the work of Bouzy &
Caveznam.

jeffrey

Juergen Kahnert wrote:

> GB> I still think GA deserves a good try.
> GB> Perhaps GA with a tree search for local tactics would work.
>
> Maybe, i don't know. Try it and share your experience, would be
> interesting.
>
> GB> I haven;t heard of anyone trying GA for go, have you ?
>
> It seems Jeffrey G. had.
> Maybe he tells us more.
>
> Juergen



--
Jeffrey Greenberg
Mgr. Adv. Dev.
Acuson Corp.
www.ultrasound.com
www.acuson.com
650-694-5422