Re: computer-go: board reduction

[Serge Boisse <boisse@xxxxxxxxxxxx>]

Dave Dyer wrote:
> 
>  Board size reduction, in the manner you are discussing it, is futile. It
> cannot possibly work. The idea of "reducing" in some way has merit, but
> it can't be just to start making the board smaller and leaving out stones.
>
That's an opinion, not a proof. futile ? It might be.. or not, depending on how
"cute" the reduction algo is. Any multi-level representation must have the
feature that some information is lost (or condensed) btw each layers. My
assumption is that go is a "spacial" game, in which spacial relationships btw
stones, chains, groups, territories.. is very important. 
I've tried a lot of ideas before coming to this one, including a program that
maintains an idea of groups of chains and groups of territories. But now I think
that purely spacial reduction can work, if the reduction process is
"intelligent" enough.
Don't be misleadeaded by the very simple reduction scheme I propose in my web
page. This one is too simple to work, of course, because it dicard stones "at
random" without considering their importance for the next stages of the game.
But this dosnt mean that an other scheme (for instance based onto feature tres)
could not work.
 
>  I've always though that a multi-level reresentation of the board
> could be useful, where the next level up from the physical stones
> would be strings of stones, next level above that wold be live and dead
> groups, etc.  I'm sure all programs contain structures like this already,
> but the trick would be to explicitly reason about higher level constructs,

I do agree with that...

> and devolve to looking at the stones only when necessary.

...but i think it will be necessary at any time...

> ....
> 
>  The point would be to reason and plan and reason top-down, and only
> pay attention to individual stones at the very end of the process.

No. I think that in oder to find the best plan, you have to perform local (for
this plan) searches.

-- 
Serge