Re: [computer-go] Data Mining

["Frank de Groot" <frank@xxxxxxxxxxxxx>]

----- Original Message ----- 
From: "Frank de Groot" <frank@xxxxxxxxxxxxxxxxx>
To: "computer-go" <computer-go@xxxxxxxxxxxxxxxxx>
Sent: Sunday, November 14, 2004 10:30 PM
Subject: Re: [computer-go] Data Mining


>
> > It's not a good excuse to say that your thing can't learn well thanks to
a
> > lack of harddisk space.

>
> Besides, I have already said that I found a solution, and that this
solution
> is not so much slower than having a 256 TB disk array, as writing 256 TB a
> few thousand times with random access takes a lot of time as well.


BTW since today, I have abandoned the "learning".

I now prefer to record hard statistical data. This only takes 24 hours.
I will find a way to use the hard data to suggest better moves than with
statistical optimization of pattern values.

The problem with automatic learning is that you draw, as it were, a function
through the pattern collection, minimizing the mean error. This works nicely
with Joseki and Fuseki and some "good shape", but after all this pattern
analysis, nothing tangible remains. There is no way to explain to the user
why a pattern is "good", there is no indication about the reliability of the
pattern's value (dependent on frequency), there is no info on how "urgent" a
pattern is etc. Just a "general statistically cross-correlated whole-board
value".

As I am focussing at the moment on a study tool and I want to present hard
data, I prefer to collect (20 bits of) hard data.

I am hopeful that the hard data, when used in neural network training, will
yield similar or better results as I obtained with "learning", with the
added advantage that the user gets real statistical info on a pattern
instead of an abstract "what's it worth" figure.


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