RE: computer-go: Neural network minimax hybrid paper

["Julian Churchill" <jjc97c@xxxxxxxxxxx>]

Hi,

 Thanks for your comments.

>
> The key issue is not whether we can implement NN or not in
> GO, but what we
> let an NN to learn, or what we knew the measure that is about
> right to make
> a decision.
>
> I think looking at 5x5 grid to decide a move will happen in 6
> steps thus
> assign a value from 1 down to 0 is too simple. How will this
> mechanism to
> deal with opening with most area open or during the fights
> most 5x5 grid
> looks similar? Picking another 6 unlikely moves to balancing
> 6 likely moves
> does not sounds right because the 0-value case already is an
> unlikely move.
>

 The idea of picking the unlikely moves was to give the network examples of
positions and moves in those positions that should not be considered
as opposed to the 'good' moves that should be consider further.
 I think the paper might not make clear that the actual program used a 9x9
network to fit in as much info as possible without slowing down the
system unreasonably. Ideally 19x19 networks would be used, but
that would take a ridiculous amount of time to train to a decent quality
and really would slow the program far too much.

> I think 5x5 grid can at most learn some local reaction, not
> even influence.
> I do not know how do you deal with corners and edges. If the
> potential move
> is still centered, the effective area is even small.
>

 In fact the idea is that that particular net _should_ be learning about
local reaction and small battles, in the future other features to the
program
will be developed to handle parts such as corners and the opening
specifically
because the 9x9 network is too generalised to give any useful response
when there is so little information on the board.
 The paper I think mentioned specially trained nets to handle the opening,
which
seemed to work quite well and also it has to be remembered that this project
does involve other things such as alpha-beta search.
 In fact in the future I hope to exchange the use of the coarse grained
"Area Finder" network, for giving the program a focus and some sort of
global strategy, with a network that is fed more abstract conceptual
knowledge
such as string info and to use this to generate a more concise and effective
strategy.

> The small 25x10x1 NN can remember about 250 patterns with
> symmetrical moves
> removed. Is not it too few for a 5x5 board area?
>
> Weimin
>

 I think in that case the use of the network is to get a general idea of
what
is a good move and not to memorise particular positions. Although this would
be
useful in some situations I think for neural networks it is better to
focus on the the learning and generalisation aspect. Pattern matching can be
done
with a database of appropriate patterns.

 Anyway thanks for the feedback, it is all very useful to me.

 Cheers,
 Jules

> ----- Original Message -----
> From: "Julian Churchill" <jjc97c@xxxxxxxxxxxxxxxxx>
> To: <computer-go@xxxxxxxxxxxxxxxxx>
> Sent: Sunday, December 02, 2001 5:52 AM
> Subject: computer-go: Neural network minimax hybrid paper
>
>
> >
> > Hi,
> >
> >  I just came back from presenting my paper at a conference
> and it seemed
> to
> > go down quite well, so I wondered if anyone here might be
> interested in
> > having a look. The title is "A new computational approach
> to Go", which
> > doesn't really give away much, so here is the abstract:
>


_________________________________________________________
Do You Yahoo!?
Get your free @yahoo.com address at http://mail.yahoo.com