[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] 2nd KGS Computer Go Tournament

Hi Don,


Hi Nick,

I would like to respectfully make these observation as an outsider coming
into computer Go.  Let's call it constructive criticism.

I have noticed there is a "computer go culture" that is embarassingly
conservative about change that might be positive but that is
different.  One of those tradition bound attitudes seems to be almost
an encouragement of programs that can't quite manage a real game or follow
simple rules.    This was probably very useful a few years ago, but it
just doesn't make sense in 2005 in my opinion.
I would like programs to do things properly. But I know how difficult it is to get it to happen.

Please read the advice I give at http://www.britgo.org/gopcres/padvice.html. Every item on that page is directed at at least one programmer who failed to follow it, when I ran the 1998 Ing Cup in London.
<rant>
Why did a certain respected programmer insist on labelling his board with a 45-degree flip relative to everyone else's, even when I told him not to?
Why did one programmer label his board including the letter I, unlike everyone else?
Why did one programmer implement a clock, but arrange for it to start running, not when he received his opponent's move, but after he had done a database update?
Why did several programs, include a world leader, playing under the Ing rules, put up an error message when a suicide move was played at them?
</rant>
This is not a "culture". There are words I could use to describe it, but "culture" is not among them.

William Shuberts resolution protocol is really simple,  simple enough to
be called elegant.

1. After 2 passes, the computers either agree or disagree about dead
   stone status.

2. If they agree, GREAT, we are done.

3. If they disagree, each program is informed via
   kgs-genmove_cleanup.

4. The play continues until both programs pass.

5. At this point, the protocol is slightly broken, but it should be
   that the game is over PERIOD after the kgs-genmove_cleanup and 2
   passes.

If the (supposedly) winning program fails to capture the opponents
dead stones, then who's problem is that?  Apparently it's everyone
elses problem except for the program itself.  Nothing elegant about
that.
Once KGS is fixed so that this works correctly, I hope more programs will start to support it. Once something like three quarters of them are supporting it, and moreover supporting it correctly and getting sensible results, I can start to lean on the other quarter to get them to support it too.

But I don't see it as important. When bots disagree on the score and don't know how to resolve it, I can cope. It is not a problem for me.

I have just recently added code to deal with seki, somewhat
imperfectly but it works most of the time.  Before this, my program
would occasionally lose won games because it would fill in those
points and get captured.

Wouldn't it be a lot "cleaner" just to give the program the win
anyway?  The fact that my program could not quite finish the game
properly shouldn't be allowed to "taint" what I cosider the "correct"
results.  I shouldn't have to be incovienced to the point of fixing
the seki code when we could simply adjusticate these positions before
it happens in order to return what really should have been the correct
results.
Before the two passes that constitute the game stop, the bots are playing Go. This is what they are meant to be competing at. If they screw it up and lose their seki groups, they take the consequences.

After the game stop, they aren't playing Go, they are scoring. This is boring and mechanical, but difficult. I do not plan to penalise them for being bad at it.

Humans, even dan-rated humans, aren't good at it either. When humans do it, they have two big advantages over bots: they are generally co-operative, both trying to attain the "right" result; and they can use language to each other. Even so, they can get it wrong, I have seen two English-speaking 5-kyus make a real mess of "resolving" a bent-four.

When humans do not co-operate after the game stop, there can be interesting consequences. The Yahoo Go server implements the game end procedure in a way of its own. Many of those who play there have learned to take advantage of this. The end of a game between two such players, with fast time limits, can be an interesting spectacle, not at all like Go. But this is not something I want to have any part of.

Ah yes, that reminds me, time limits. I believe that the official Chinese rules fail to say whether clocks are used after the game stop. The Japanese rules certainly don't say. So we have two problems here. If we don't know what the rules are about applying time limits to the game end, how can we enforce those rules? And, if there are no time limits, then the bots can take as long as they like establishing the status of a group, and still be doing it when the next round starts.

Ok, I'm being very sarcastic to make a point, forgive me.  No harm or
disrespect intended.  But I feel it's no different with dead stone
resolution, it's nothing like rocket science.  I believe computer go
programmers are smart people and can handle this properly.  It's a
simple thing
I believe computer go programmers are smart people. I have doubts about whether they can handle this properly, I suspect it is more difficult than you think. And I am sure that some won't handle it properly until a lot of pressure is applied.

and it should be considered as essential as knowing any
of the other rules.
!! Knowing rules it not essential.


. # # O . . O . . Full 9x9 board.
# . # O . . O . . Japanese rules.
# # # O O O O O O There are no prisoners.
O O O # # # # # . Each player has just passed.
. . O # . . . # # Neither player feels any need to make any further
. . O # . . . . . move.
. . O # . . . . .
O O O # . . . # . What is the score?
. . O # . . . . .



. # # O . . O . . Full 9x9 board.
# . # O . . O . O Japanese rules.
# # # O O O O O # There are no prisoners.
O O O # # # # # . Each player has just passed.
. . O # . . . . # Neither player feels any need to make any further
. . O # . . . . . move.
. . O # . . . . .
O O O # . . . # . What is the score?
. . O # . . . . .


I doubt that one player in a hundred, at around shodan level, can answer these questions correctly. But I would not want to penalise them for it.

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