RE: [computer-go] Chess programs versus go programs

["Mark Boon" <tesuji@xxxxxxxxx>]

> -----Original Message-----
> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of Eric Boesch
> Sent: Wednesday, December 08, 2004 18:31
> To: computer-go@xxxxxxxxxxxxxxxxx
> Subject: Re: [computer-go] Chess programs versus go programs
>
>
> It sounds like you are suggesting computational game theory (CGT)
> calculations.  Does anybody's program do that?  (I'm guessing that most
> programs have some kind of code to deal with sente vs. gote issues, but
> that's not the same as full-generality CGT.)
>
>     http://senseis.xmp.net/?path=Endgame&page=GoNumbers
>
> Even a brain-dead CGT calculator might be of some use, since people are
> terrible at math; such a program could probably solve most of the
> problems
> on the above page in microseconds, for example.  But as you say, a CGT
> calculator can benefit from (and for more complicated problems, it would
> require) the same move-pruning and evaluation hints that speed up
> ordinary
> local search.  If a CGT calculator used hooks to existing
> evaluation code,
> it would not add too much complexity to an existing program, I think.

I've made a serious attempt on this in the past, but didn't have much
success. The endgame positions in Berlekamp's examples are very simple and
not much use in real play. I tried to make a module that would compute the
'value' of simple local end-game positions like the following:

          |...XXO..   |.XOOO..
|XXXOOO.  |..dXOO..   |..XXO..
|.......  |..bac...   |.......
|------   |--------   |-------

After a few weeks the results were still very disappointing, basically
because even on these simple positions it took forever to get a result. Lack
of time made me put this on hold and I never got back to it. That was almost
a decade ago though. Since humans can compute proper values for these
(although often they only think so), I figured a computer should be able to
do it too. But I guess it was a little harder than I thought.

Take the second example. The hane-and-connection on the 1st line (a-d) is
white's prerogative. The reason is that white's follow up in case black
doesn't answer is bigger than the value of the move that black could play.
But that is a chicken-and-egg problem, since you're trying to compute the
value of that move as it's part of the 'value' of the position. What I tried
was an iterative process to get a clamp on the value-range up to the point
it could deduce 'a' is always sente for white. But this iterative process
too often just ended up in a runaway calculation.

I did notice though that there are almost no sente-moves in any of the
positions on which I saw CGT applied, so maybe there's a difficulty there.
Or maybe there have been develoments in the meantime that I missed :)


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