Re: Fuseki and Joseki Database with Neural Network

[Jeffrey Greenberg <jeffg@xxxxxxxxxx>]

Greetings,
I've spent quite some time on using GP for GO, so I understand your
difficulties.

You may very well have to give up on the idea of a quick evaluation function,
but that depends on the problem you are trying to solve.  If the problem
requires a very high-level of Go knowledge, then evaluation times will be
significant given the current technology.  But, there is quite a bit of
literature about doing GA and GP with this condition, such as where the
evaluation is done by a person...  If you want to use a simple function that
determines whether a move is or isn't in the joseki database, then you've got a
quick evaluation function but perhaps too little smarts to evolve anything of
substance 'before the sun explodes'.

Hybridizing a neural net with Go & Gp is a good compromise, since the ANN seems
to acquire complexity a bit more efficiently than Go or Gp at this point in
time.  Other GA/GP hybrids may also be appropriate.

Another problem you are facing is: you need to create a joseki individual that
is exactly correct as opposed to having one that is correct on average;
otherwise, it can be exploited.  In other words, robustness.  How do you propose
to train?  How do you propose to capture sequences of moves?  How about deep
ladders which beg for loops and recursion?

The joseki problem is a good one.  To the extent that joseki can be represented
as simple local trees, why bother trying to "learn" them when they can simply be
entered?  On the other hand joseki books tend to ignore the global situation and
thus are simplifications in their own right.  Trying to synthesize joseki in a
global context would be challenging.

jeffrey greenberg
ps: some of my work is at http://www.cris.com/~jgberg

Juergen Kahnert wrote:

> Ok, i exaggerates with the used time span for the evaluation function.
> ;-)
>
> But i think, the time will not be the main problem. Rather to find a
> really good evaluation function which will work for genetic algorithm.
> I mean, it is impossible to find one.
> Why?
> For GA we need a function with a steady course (i don't know if it is
> the correct mathematical expression for a function without any 'jumps').
> We need a function with minima and maxima. We are able to avoid to end
> in a local maxima by vary the step width, but this requires such a
> steady function.
> And thats the problem. I think we can't find such a function because in
> "go" it's possible a stone placed on a field maybe bad but the field
> next to it could be very good. It is not a little bit better or
> something like that, it could be like day and night. That produce a
> 'jump' in the function.
> Now a generation near the best place would die because it's current
> place is very bad. The solution is to make steps with one field, but
> this means we have to try all the possible fields and this means the
> advantage of GA is lost.
>
> Maybe if have a wrong feeling with this, because i'm not very good in
> the opening. But if there exists a function i'm not able to find it and
> if someone is able to find such a function i think he or she will not
> use GA because it's no more necessary... ;-)
>
> So i come to the conclusion neither genetic algorithm nor neural network
> is the correct way to do fuseki or joseki.
> Do you accept with me?
>
>         Juergen
>
> PS: I have not heard of the GA software "SAMUEL".
>
> Gary Boos wrote:
> >
> > Ahh.  Good comment - I think you are normally correct.
> > But isn't it true that if you need some level of accuracy (better
> > than a NN would tend to give) and you can let you machine run
> > for 10 -20 hours to produce the "solution" (what is the proper term ?)
> > using the GA, wouldn't GA be better than NN?
> > You suggested  "before the sun explodes" time span.  Is it that
> > difficult to write an eval' function (and somehow limit the number of
> > generations or mutations with each new generation) that a "solution"
> > after 10-20 hours is feasible?
> >
> > Please comment (though I would think this is covered in a basic
> > commentary on the subject is a textbook).  You could just answer
> > - give it a try and see how hard it is.
> >
> > Thanks, Gary
> >
> > BTW have you heard of the GA software "SAMUEL" ?



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