Re: Fuseki and Joseki Database with Neural Network

[smg6@xxxxxxxxxxx (S. Max Golem)]

>>  I'm not an expert on ANN. But I'm not sure joseki is best solved with
>>ANN.

Neither am I, but I had to work with them in three grad courses several
years back. Details are a bit fuzzy, but I hope some of my general
knowledge/intuition is still accurate. Based on that, I have to agree that
joseki is not a likely application for pure neural nets.

>> Because my experience with ANN is that it is not exact. But ANN can "guess"
>> where it is not sure. 

I think this ability to interpolate between known cases is a problem here. 
First, the moves in joseki often are discrete and have to be made with care
-- two possibilities leading to totally different variations may be a few
spaces apart, but if you take the middle point (as generalizing NN's are
apt to do), you get a silly move that may cause utter ruin.  The same if
the NN's guess is one point off.  This interpolation can be very powerful
in object recognition from video input, but probably a drawback in joseki
recognition. 

Second, the NN's ability to generalize is a two-edged sword -- you can't be
sure what is being generalized since the weights inside the NN are
invisible and by no means unique.  For example, the US military in the
early days of NNs wanted to develop a system that would scan picture data
from aerial reconnaissance photos automatically to detect the presence of
tanks in the landscape.  After training their NN extensively and getting a
high correlation (70 - 80% correct, I think), the developers felt ready for
a demonstration to the top brass, who brought their own photos for
analysis. 

Anyway, the computer failed miserably, scoring little better than random
guessing.  It tended to miss tanks that were in bright sunlight and quite
obvious. 

Later analysis showed that the NN was primarily choosing the photos based
on how overcast the day was, the presence of tanks being a secondary
consideration.

Obviously the problem lay in training the net with non-random data, but the
real point is that -- because a NN is a "black box -- you can never be
certain what you're training it to recognize. 

Another big problem is that of convergence. The bigger a NN is, the longer
it takes to train, even with perfect data. A NN large enough to do anything
with joseki is likely to take an immense length of time to train. If you
want a useful result within one human lifetime, the NN has to be designed
very carefully and even then there's no guarantee (getting a NN to converge
during training is a major consideration in NN design, and this is another
aspect in its infancy -- and the math involved can be truly horrendous).

>>So maybe ANN is good for choosing fuseki and maybe
>> choose joseki depending on the fuseki. But to store a joseki sequence I
>> think normal datamodels is better than ANN.

Agreed.  Since finding the correct josekis for a given position has such a
high tactical consideration and pure joseki theory is so immature, I think
a large joseki database is the best way to go for the foreseeable future. 
Of course a hybrid model using NNs to evaluate thickness and so forth would
probably provide useful input in joseki selection. 

> >To explain why datamodels is better I think the following is a good 
>way of thinking.  

>At each step in a sequence of forced moves in a joseki, there are >many
alternative moves that look good as local tactical shapes, 
>but nethertheless the joseki only allows for one correct move - all 
>the other alternatives are mistakes.  

On the other hand, there are often slight variations that are only
minimally inferior, and the program has to know when an how to punish
departures from joseki. Often, you have to just leave it and concentrate on
the weakness in the middle game. Or maybe there is no real refutation --
the offender simple negates some aji or takes a point less. How computers
will deal with this will be interesting.

Oops! I'm about to be cut off. Enough for now anyway.

Bye,

Max.