RE: [computer-go] Pattern matching - rectification & update

["Mark Boon" <tesuji@xxxxxxxxx>]

> -----Original Message-----
> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of Frank de Groot
> Sent: Saturday, November 20, 2004 11:16
> To: computer-go
> Subject: Re: [computer-go] Pattern matching - rectification & update
>
>
> From: "Darren Cook" <darren@xxxxxxxxxxxxxxxxx>
> Subject: Re: [computer-go] Pattern matching - rectification & update
>
>
> >
> > What is this 90% of? 90% in a life and death problem test? Or
> looking at a
> > board and correctly saying alive, dead or unsettled(*) for 90% of the
> chains
> > on the board?
>
>
> I meant, there are now complex tactical problems that take too long to
> solve, because the solving code wants to prove a solution.
>
> I was wondering, would it be possible to be much faster and only be "quite
> sure" that it is solved, and would that still, in a Go program, lead to
> reasonable results?

One more time: you can do this with proof-number search.

I've found that given a search of a few hundred nodes gives a pretty good
first idea whether a group is weak or not. Provided you have a reasonable
evaluation of course. The weak ones can then be investigated further, while
you consider the others to be 'alive' or 'strong'. You don't need to search
until proof, after each expansion you have a new estimate how likely the
proof and how likely the disproof is at that point. You can have it stop at
any 'likelyhood' you want. There's no guarantee how long it will have to
search until that likelyhood is reached of course, it can easily result in a
runaway search. To control this potential runaway effect will be one of the
main challenges.

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