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

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



I've noticed some of the stronger programs can't read out things like seki all the time. Which is reasonable, because if they were perfect at reading they would be 9p+ and this bulletin board wouldn't be needed any more. But this is common enough to make human intervention in scoring necessary - you can't play out a seki!

Paul


----- Original Message -----
From: "John Davies" <john@xxxxxxxxxxxxxxxxx>
To: "computer-go" <computer-go@xxxxxxxxxxxxxxxxx>
Subject: Re: [computer-go]  2nd KGS Computer Go Tournament
Date: Mon, 09 May 2005 13:11:45 +0100

> 
> At risk of being stepped down on ;-)
> 
> I hadn't thought of the contention case Karol outlines below which makes
> my bullish comments on testing seem a bit grumpy; maybe there is a gap
> which could be closed in the protocol there.
> 
> When I struggled with getting Dumbbot to assess dead groups properly I
> found it was much harder than I'd expected (especially with some of the
> more advanced bots like GNUGo leaving things it knew it could kill if it
> had to, rather than making things obvious to lesser bots).
> 
> My solution in the end was to spool out the final position in SGF to a
> local GNUGo process, ask it to "final_status_list dead" and pass on the
> result.  I plan to take this assessment job on myself eventually but for
> now it doesn't improved DB's results (so isn't cheating) but does make
> sure the live/dead status is very good at the end. Result: a happy
> tournament organiser :-)
> 
> This trick isn't tricky ;-) and could be suggested for any bots that
> don't do this stage well.  I'm surprised the winning bot (obviously with
> a strong sense of what's alive and dead during the game) was lacking in
> this regard.
> 
> BTW I don't count this as a 'hybrid' program so PLEASE don't start that
> thread going again!
> 
> HTH
> 
> JD
> 
> 
> On Mon, 2005-05-09 at 11:17 +0200, Karol Golab wrote:
> > On Mon, May 09, 2005 at 10:05:03AM +0100, Nick Wedd wrote:
> > > If the problem were genuine disagreement, I would agree with John.  But > it 
> > isn't.  Some bots have game-end-handling code which is much worse at > assessing 
> > status than their actual playing code.  For instance, LeGoBot > plays at around 
> > 20k, but in its first-round game, it claimed at the > game-end that its 
> > one-liberty black group, entirely surrounded by > unkillable white groups, was 
> > alive.  LeGoBot's playing engine knew it > was dead, the error was merely in its 
> > scoring engine.
> >
> >   I think that the actual problem was more on KGS side - as far
> > as I know LeGoBot did not return any status at all (it does not
> > implement final_status_list command from what its author said).
> > The problem occured because tlsBot returned the status and LeGoBot not.
> > I would assume that the KGS engine interprets this as 'all stones
> > live' which led to a disagreement.
> >
> >   Cheers,
> >   Karol
> >
> 
> _______________________________________________
> computer-go mailing list
> computer-go@xxxxxxxxxxxxxxxxx
> http://www.computer-go.org/mailman/listinfo/computer-go/

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