Re: computer-go: Discarding the rubbish moves

[Heikki Levanto <heikki@xxxxxx>]

On Wed, Jan 15, 2003 at 11:27:21AM -0600, Richard Brown wrote:

> > Heikki Levanto wrote:
> > > Anyway, in any given go position, there are a small number of
> > > reasonable moves that can be considered, and a great number of
> > > totally stupid moves that no human would even consider.
> 
> Heikki has made a very important observation.  I hereby dub it
> "Levanto's Observation", in honor of the observer.

Oh, thank you! I am not at all sure if this is such a great observation, nor
if someone else may not have voiced it before. 
 
> > > Any go
> > > program must be able to discard the 90% of rubbish moves quickly.
> 
> Yes, this is extremely important.  Strong programs of the future will
> use such a razor in an early phase. I believe that the actual percentage
> will be higher.  I'll go out on a limb and predict 97%.

Well, one problem is that there are a lot of moves that are locally
playable, but make no sense in the global position. For example, playing
small endgame moves at the end of a joseki.

Another problem is that there are moves that look totally stupid, and 99% of
the time are not even worth considering - but occasionally they do get
played, because of a combination of various details (forces the opponent to
play a slightly worse shape reply, while enabling a strange follow-up that
breaks a potential ladder in some far-away fight, which fight is then more
likely to take a development that ... which will enable me to play a move
that corrects my ugly shape). Good luck extracting and encoding such
features!
 
> But what makes one classifier better than another?  This is measurable
> by how often it is wrong; that is, how often it incorrectly vetos a
> move,
> only to see a pro play it.  (Darren's .1%)  Somewhat harder to tell, is
> how often a classifier didn't veto a move that it should have.  (Don't
> we have to ask every pro?)

It should be sufficient to ask just one pro to list all moves worth
considering. Unfortunately such information is not readily available. We
know which move the pro played, but we can not safely assume that any move
he did not play was immediately out of the question - he may well have
considered it, found it reasonable, but chosen the game move instead for
some minor reason, like knowing the opponents style and noting his condition
on the playing day - or purely by random.
 
> I'm still trying to get my head around the fact that it is possible to
> extract an _infinite_ amount of information from a finite amount of
> data.

Scary thought. Here is another: There will always be an infinite amount of
features that were present when the data was created, but that did not get
captured in the data.
 
> "Go is the damnedest thing I've ever seen."

I guess we can agree on that!

-H

-- 
Heikki Levanto  LSD - Levanto Software Development   <heikki@xxxxxxxxxxxxxxxxx>