Re: Fuseki and Joseki Database with Neural Network

[Juergen Kahnert <Kahnert@xxxxxxxxxxxxxxxx>]

DM> I thought the idea would be to use the presence/absence of the

DM> suggested move in the traditional joseki database as the

DM> objective function.



Correct. This is a quick evaluation function.

Now, i have some trouble to use this "new" information, because i can't

imagine the improvement of this knowledge - it's not really new. So, why

not take this move directly out of the database?

Sorry, i couldn't find a way to make a use of it.

How to find a good alternative sequence - which is not a part of the

database - if all move not contained in the database are bad?



DM> But you certainly don't need a continuous objective function

DM> for GA. You don't need a continuous "target" function for

DM> neural nets for that matter.



I can't follow you, sorry.

To get the fitness of a generation I need a continuous function. ;-)

Why to let alive a good geneartion if the periphery fields are all bad?

I can't approximate step by step to the maxima while having "jumps" in

the evaluation function...
A target function in a neural network is for one neuron, not to evaluate
the whole output. To evaluate the output of a NN you have a look at it
an decide wheter it's good or bad. If bad you have to train the NN
again.
While using GA you know if something is good, you want to improve it
(not to get some results, no, you want to get a better one).



The representation problem of a joseki sequence for GA is another
problem.
I have not thought about it as long as i'm not sure that the problem is
solvable.
But you are right, it could be very tricky. :)

DM> Imho, the direction that people should be going in, with regard to

DM> machine learning approaches to computer go, is to:

DM>

DM> 1) think hard about improving the representation of the problem that
the

DM> system works off of - don't just feed it the raw board position and
hope

DM> it will figure out about eyes, life, connectivity, etc.

That's right. Only by having some positions the computer isn't able to
realize the connection between the stones...


DM> 2) rather than just trying to solve go, attack subproblems: move

DM> ordering during tactical search, evaluating whether a point will end
up

DM> as territory, whether groups will end up alive or dead, whether

DM> releationships are severable, whether a stone is tactically
capturable,

DM> etc.

Yes, i think so, too.
One of the biggest problems i see is, how to represents the knowledge of
an artificial intelligence?
It's one thing to feed the computer with static algorithm to solve all
these problems and a completly other thing to let it learn...
How to collect the knowledge and how to store it for quick access?
Making a function which is able to avoid to run into a ladder is much
more easier than making a function which will learn to see and avoid it
or use it...

Juergen