Re: [computer-go] I know we disagree, but I choose to donothing about it.

[Tom Cooper <t.cooper@xxxxxxxxxxxxxx>]

At 04:58 23/07/2005, Don wrote:

> The proposal is for automated game scoring and it's not this dramatic.
>
> The year is 2005 and 2 computers still cannot even play a game of Go without
> getting a human involved in the scoring process.  Even with Chinese 
> scoring.
>
> That's what this is about.   It was  ok for the early pioneers of Computer Go
> but we should be way past that by now.    I know that in general computer go
> is very conservative and leery of change,  but this should be a no-brainer in
> my opinion.
I disagree with this point of view.  I think it is nice to have the required
entry standard into these tournaments as low as possible, it allows more people
to enter, and to enter earlier in the development of their programs.  Also, 
a lot
of programs don't build on earlier ones, so I don't think the year is 
relevant.  Some
programs work by pattern recognition, neural nets, genetic algorithms 
etc.  I don't
think it is realistic to expect that this type of program could implement a 
game
end protocol without adding a lot of a traditional-type go program; 
especially if the
data they learn from does not play out to the end of the tromp-taylor game.

David wrote:

> I see the details of the suggested protocol as quite different from the way
> I have always played Go. I have taught hundreds of people to play Go,
> and in the first few games this virtually always occurs because new players
> need to learn what living and dead groups look like. I know full well what
> stones are alive and which are dead, but they do not. So they say "why
> aren't these alive?" and I say "finish the shape so they have 2 eyes." So
> they make a move. In most cases the stones are still quite dead, so I pass
> again and tell them the stones are still dead and to keep trying. Usually
> after a few of their moves I will have to play a single stone to assure that
> the stones in question stay dead. If they pass at after one of my passes 
they
> do not get to claim all the stones on the board are alive. If they say 
"would
> you please play first so that I can see what you mean" I will, 
explaining that
> if this were not a teaching game I would not be required to play.

I disagree with this too :).  While it might be sensible to teach this way, it
is not the way a competitive game of go works.  Although the standard of some
programs is at or below that of a complete beginner, in a tournament, those 
programs
are trying to win, and not to teach each other.  In a competitive game, it 
is not
enough to say 'that group is dead' or 'this group is alive': you have to be 
able to
prove it (although experienced humans normally agree on status and don't 
have to bother).
I particularly disagree with the last two sentences I have quoted.  Firstly
they have as much right to claim their stones are alive as you have to 
claim that they
are not.  Secondly, if you disagree about the status of their stones, then 
in the
confirmation stage, you WOULD be required to play, and to take them off the 
board,
in the sense that if you didn't, they would be counted as alive.

I don't think there is a simple algorithm to decide the score according to 
common sense
(as much as there even is a common sense here) of a game in which the 
players disagree
about the status of groups, but won't play to remove all dead groups.  I 
think Nick is
inevitably going to face tougher and tougher judgement calls.  So far, from 
what I have
seen, he has erred on the side of common sense (quite rightly I feel) and 
the competitors
involved have been gracious in accepting his decisions, but there is no 
guarantee that
this will continue to be the case, and it would be nice if the requirement 
for human
intervention could be removed.

Would it be feasible or desirable to set things up in such a way that gnugo 
plays out a
kgs-genmove_cleanup for programs that don't implement it themselves?

Cheers,
Tom.




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