Re: [computer-go] ANNs as potentially useful in the computergoproblem...

[David Fotland <fotland@xxxxxxxxxxxxxxx>]

I think machine learning will be essential to strong go programs.  Many Faces
of Go doesn't use machine learning at all, unless you count memorizing opening
positions from pro games and its own prior games.  But the rate of 
improvement for
hand-tuned go evaluations is way to slow to get strong in my lifetime.

David

At 08:47 AM 1/10/2004 -0500, you wrote:

> Jim, I'm on your side.  I'm not berating NN's like Vincent is, I think
> Neural Nets  need to be explored a  lot more.  In my  opinion what has
> been done so far is just scratching the surface.
>
> I read  the book in full  and I loved it,  but I stand  by my original
> statement, the program  was not tested against a  "real" program and I
> doubt it would  stand up against one, but we really  won't know how it
> would do.
>
> Chinnok was only  impressive when it played at  full strength with all
> the hardware.   Even if  it had been  against a full  strength Chinook
> (and  it wasn't) a  single win  out of  1 game  tells you  very little
> except that it is likely to be within a few hundred rating points of
>
> The  points Vincent  makes are  based on  the fact  that  NN's haven't
> appeared (as a  major part) of a top notch GO  or Chess program.  It's
> pretty easy to be negative when this is the case, harder to be open.
>
> - Don
>
>
>
>    From: "Jim O'Flaherty, Jr." <jim_oflaherty_jr@xxxxxxxxxxxxxxxxx>
>    Date: Fri, 9 Jan 2004 23:46:53 -0600
>
>
>    Don,
>
>    You might want to read his book.
>
>    His final chapter was taking the final output from Blondie24 and 
> playing it
>    against the world's BEST Checker playing program, Chinnok.  And the
>    Blondie24 840th generation winner beat Chinnok in a single 
> game.  Granted it
>    was not a series of games.  That was not Fogel's goal.  It was to
>    demonstrate he could produce an expert level player and that it could beat
>    Chinook.  One game was enough TO SATISFY HIS EXPERIMENT.  As far as I am
>    concerned, Fogel clearly demonstrated that ANNs are "adequate for the 
> task",
>    to quote from your post below.
>
>    And if you had read his book, you would be aware of how he used large
>    numbers of games against humans, some of which were at the expert 
> level, to
>    determine the level of his resultant computationally generated players.
>    Fogel did his experiment quite nicely, I think.  And the fact that I am
>    successfully reproducing it gives me more confidence in utilizing ANNs and
>    GAs to pursue my Computer Go goals.
>
>    Again, I am not saying ANNs and GAs are "the answer".  I am merely saying
>    that Vincent's complete disregard of them as inadequate or incapable seems
>    pretty unsupportable given the evidence presented by Fogel AND 
> REPRODUCED BY
>    MY OWN EFFORTS.  I personally am speculating they both will be significant
>    part of the answer that generates a don level Computer Go player.  *When*
>    that will be, that is another matter entirely.
>
>    I am fine with Vincent's lack of confidence in ANNs.  That is his 
> issue that
>    he is cutting off a whole branch of options in his search space to 
> generate
>    a good Computer Go player.  It is no skin off my nose.  I will be 
> generating
>    my own experiments and my own confidences and very unlikely to cut of any
>    path of exploration based on unsupported and likely unsupportable 
> assertions
>    made by Vincvent or anyone else.
>
>    I will, however, be listening VERY CLOSELY to those who have generated the
>    better Computer Go players, like David Fotland and company.  I suspect 
> their
>    experience will be invaluable in considering all of the different ways 
> to go
>    about encoding essential Go information values regardless of the 
> "knowledge
>    search" implementation methods.
>
>
>    Jim O'Flaherty
>
>
>    ----- Original Message -----
>    From: "Don Dailey" <drd@xxxxxxxxxxxxxxxxx>
>    To: <computer-go@xxxxxxxxxxxxxxxxx>
>    Cc: <computer-go@xxxxxxxxxxxxxxxxx>
>    Sent: Friday, January 09, 2004 5:01 PM
>    Subject: Re: [computer-go] ANNs as potentially useful in the computer
>    goproblem...
>
>
>    >
>    >
>    > Jim,
>    >
>    > Blondie24  demonstrated  pattern   learning,  exactly  what  NN's  are
>    > supposed  to  do.  The  program  improved as  it  was  trained and  it
>    > improved significantly.
>    >
>    > However  there was  no comparison  between Blondie  and  actual strong
>    > checkers programs.  I don't think Vincent is claiming that NN's do not
>    > learn, only that he believes they are not adequate for the task.
>    >
>    > So I don't think Blondie proves (or disproves) anything.
>    >
>    > However,  I didn't  notice anything  useful in  Vincents  comments, he
>    > merely made an  assertion that no one can prove.   I can't imagine him
>    > ever finding  a proof  for this,  but I can  imagine someone  at least
>    > having a chance to refute it by showing a strong program.
>    >
>    > But it has to be stronger or  at least very close to the best attempts
>    > without Neural Networks.  I think  this is very hard and makes Vincent
>    > feel safe.
>    >
>    > - Don
>
>
>    _______________________________________________
>    computer-go mailing list
>    computer-go@xxxxxxxxxxxxxxxxx
>    http://computer-go.org/mailman/listinfo/computer-go
>
> _______________________________________________
> computer-go mailing list
> computer-go@xxxxxxxxxxxxxxxxx
> http://computer-go.org/mailman/listinfo/computer-go

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