Re: Neural Net Computing

[Hyoungsoo Yoon <hyoung@xxxxxxxxxxxxxx>]

I'm also very interested in neural network approach,
and I believe there's real potential here considering
the importance of pattern recognition in go.

However, go is also very knowledge-intensive and
previous attempts haven't been very sucessful, IMHO.
If you like the references, one of the most "succesful"
work was done by Schraudolph, Dayan, and Sejnowski.
The work was based on learning technique called TD(lamda),
which had been very successfully used for backgammon,
and to some limted extent for chess.
It was published in NIPS'93. pp817-824.

(I don't want to discourage anybody, but if you don't know
what TD(lamda) means, for example, I wouldn't recommend
you to try neural net approaches. This is simply a reflection
of my skepticism toword using nn methods in go programming,
and if  you don't consider yourself as "expert" in nn, your
time will be better served by trying more 'conventional' techniques.)

I know there have also been hybrid approaches such as
genetic algorithm + nn, expert system + nn, etc. But I dont
have any references. If anybody has extensive references
on these methods, I'd really appreciate it.

Hyoungsoo

----- Original Message -----
From: Tim Boldt <tim_boldt@xxxxxxxxxxxxxxxxx>
To: <computer-go@xxxxxxxxxxxxxxxxx>
Sent: Wednesday, November 25, 1998 4:28 PM
Subject: Neural Net Computing


>It seems like neural network computing would be applicable to Go (at
>least for higher-level strategy decisions).  Have there been any
>attempts to do so, did any of these attempts show promise, and if so
>where could I go to find more?
>
>Thanks,
>
>Tim Boldt
>tim_boldt@xxxxxxxxxxxxxxxxx
>
>