RE: computer-go: Board evaluation by counting...

["Mark Boon" <tesuji@xxxxxxxxx>]

> -----Original Message-----
> From: owner-computer-go@xxxxxxxxxxxxxxxxx
> [mailto:owner-computer-go@xxxxxxxxxxxxxxxxx]On Behalf Of Julian
> Churchill
> Sent: Friday, September 21, 2001 1:50 AM
> To: computer-go@xxxxxxxxxxxxxxxxx
> Subject: RE: computer-go: Board evaluation by counting...
>
>
> >
> >
> > That sounds more like the NN approach discussed here before.
> > I have severe
> > doubts this will actually work.
> >
>
>  Well it can only be tried, if everyone just said that's not going to work
> without attempting to experiment with things a bit and have a go at
> contributing some useful experiences to a topic then we wouldn't get
> anywhere. Even negative experiences are useful since they rule out methods
> to try and often suggest a good deal of alternative approaches that could
> yield successful results.
>

That's true, it can only be tried. In this news-group new people who want to
try this pop up with a steady regularity. There must have been a dozen or so
people who wrote they were going to attempt this seriously. We never heard
of any of them or their project again. So, yes, it would be nice to hear
about the failures as well, so as to prevent other people to waste their
time trying the same thing.

> >
> > It is not at all an evaluation function, but an influence function.
> > Influence is just a small component of an evaluation function in Go.
> >
>
>  But surely an influence function can be used to estimate territory and
> hence be a good measure of the relative position of the players at that
> stage, exactly the criteria for an evaluation function? Influence may be
> just a part of a more complex evaluation function but I see no
> reason why a
> neural network that was trained to be used as an evaluation function,
> perhaps using temporal difference methods, would not be able to
> assume some
> of the other components of a more complex eval. function. As I
> said before,
> we don't know what is possible until it is attempted and the
> neural network
> area is a relatively new and rapidly expanding subject where new
> methods are
> being discovered/invented all the time.
>

No, that's a mistake in thinking. The component that has the greatest effect
on the evaluation and on the outcome of the game is the life-and-death of
groups and to a lesser extent the strength/weakness of groups. Influence
only has a small effect on the outcome and just on its own is a bad measure
of the relative position.

> > What is the similarity between Go and Backgammon? Is there
> > any at all that
> > would suggest that which works for one is going to be
> > remotely useful for
> > the other?
> >
> > In general I don't believe in the 'magical' approaches where
> > one single
> > solution will give a Go playing program. Go is too complex.
> > Feeding pro
> > games to a learning algorithm and hope it will learn to play
> > Go is like
> > feeding Shakespeare to a monkey and hope it will learn English. (The
> > comparison is actually an insult to the monkey, which is
> > capable of learning
> > a great deal more than any known learning algorithm.)
> >
>
>  I'm not suggesting an all-powerful network should be created,
> that is quite
> clearly a highly implausible solution. Even so I think you
> underestimate the
> possiblities that learning methods present, after all humans aren't
> programmed with hard-coded rules before they can play a game,
> they learn the
> rules and benefit from experience over a lifetime. If a human was
> only given
> the rules but never allowed to correct their mistakes from game
> to game, or
> even from move to move, no improvement would ever be made, so I firmly
> believe learning methods have a very important role to play in all AI
> applications not just game playing.
>
> What may be possible is to use neural networks in conjunction with other
> methods to gain the advantages of a program that can learn. As I
> see it one
> of the main problems with Go programs at the moment is that after a player

> has had a few games the weaknesses and strengths of the program
> are readily
> identifiable. If the program had a learning element it would be able to
> adapt to cope with a human trying to exploit it's weaknesses just as any
> human would quickly learn to do.
>
>  Cheers,
>  Julian Churchill
>

I'm not underestimating the possibilities of machine-learning at all. In
fact this was one of the main conclusions of a talk I gave at the 1991 In
Cup that if Go-programs were to significantly improve they would need some
sort of self-learning mechanism because the amount of knowledge that can be
hard-coded by a human beforehand is too limited.

I agree that NN or other ways of machine-learning in conjunction with other
methods could be successful, but you never pointed out what those 'other
methods' are as it seems to me the way you described it it's not much short
from a NN that tries to learn the complete game. Therefore, whenever someone
writes here about feeding pro-games to an NN and hopes to learn something
significant from it, I try to temper the enthusiasm and suggest trying
something slightly less ambitious first.

So I'd like to see someone try to make a connect-four program based on a NN
to see if a network like that can learn to read ahead in an automated way.
Or an influence functions that doesn't have the drawbacks of the current
static ones. And then in a few decades, when computers are a thousand times
faster than they are today, maybe some of these methods can be put together
to build a more complicated and sophisticated learning machine.

    Mark Boon