RE: [computer-go] Pattern matching - example play

["Mark Boon" <tesuji@xxxxxxxxx>]

> -----Original Message-----
> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of Frank de Groot
> Sent: Wednesday, December 01, 2004 12:03
> To: computer-go
> Subject: Re: [computer-go] Pattern matching - example play
>
>
> I wrote:
>
> >It's trivial to demonstrate that a programmer's learned patterns are
> >inferior to directly learned patterns, as I have convinced myself I have
> >done.
>
> To clarify: I did not mean that a pattern-system outperforms a Go-playing
> programmer when the Go playing programmer plays Go, I mean that when a Go
> playing programmer puts in patterns from his memory into a pattern-system,
> the result is worse than when those patterns are learnt from games.
>
> If anyone disagrees, I would like to see their pattern system
> predict 46% of
> moves in a pro game.

You have to remember that predicting pro moves is not the goal of
pattern-systems in a Go playing program, so I'm not sure it's a valid
test-case.

There are two sides to pattern-matching that are important in a Go playing
program. You want the pattern-system to be able to find important moves, and
you want those patterns to contain a relative value of urgency so you can
order them. The tweaking of the urgency-values is tricky business and I can
see that maybe a program like yours would be better capable of doing that.

The first part is more important though. Take for example the pattern of a
simple, direct cut.

    X.O
     OX (extra conditions either X's cannot be captured after playing the
cut.)

Having this hand-coded pattern in my database makes sure that my program
will *always*, 100% guaranteed, find te cutting move. Now even if your
system predicts 99% of pro moves, but occasionally misses a cut... then I
think you have a problem.

This is just one example though. There are many others almost equally
fundamental. They are so fundamental that I don't dare to risk it leaving it
to some pattern-harvester and be comfortable it will contain all of the
required patterns. You can't have your groups be cut, just because the
particular stone-formation didn't happen to match any of your harvested
patterns.

So I suppose it's also a matter of confidence. If you have confidence that
your system will also include all the really fundamental shapes it's maybe a
different story. However, simply 'believing' based on a good percentage of
predicting pro moves is not good enough for me. Would there be a way to test
this? Could we make a set of patterns like the one above that I think are so
fundamental you can't afford to miss one and see if your system contains
them? Or would it be too much work...?


>
> Now, if this is valid for "shape" (learned outperforms manually
> entered) why
> would it not be valid for eye shapes for example?

Do you think you can determine eye-space purely based on patterns?

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