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

Re: [computer-go] Mass estimation of game results?



Adrian,

For your purpose,  I think it makes sense to let GnuGo play out ALL the games. 

Start from the move just before the double pass,  let GnuGo finish the game 
and score it, regardless of the actual results.   Just pretend the opponents 
played this way.   There will be little error.  For a sanity check you can of 
course keep statistics on the games that ended differently (by GnuGo) than 
the score of the game indicated.

- Don


On Sunday 21 August 2005 11:43 am, Adrian Petrescu wrote:
> Thank you! That will work :)
>
> I know that I face a rather large risk of being wrong, but my rationale
> is that if the game is so one-sided that the player decided to resign,
> it will probably be a dramatic enough position for GnuGo to pick it up.
> Besides, all it has to do in those cases is choose between B or W, I
> don't care about the score difference if they resigned.
>
> The only alternative is to hand-check 30,000 SGF's ... I don't plan on
> doing that =P
>
> Thanks again for the idea!
>
> Arend Bayer wrote:
> >On Sun, 21 Aug 2005, Adrian Petrescu wrote:
> >>A friend and I are in the process of compiling a fairly large database of
> >>amateur games from Go Servers, like IGS, KGS, etc. We recently came into
> >> a collection of about 30,000 IGS games but they have a slight problem --
> >> they are all missing an RE[] tag, so they are heavily decreased in value
> >> for a database. What I was thinking of doing is writing a small script
> >> to run all of their final positions through GnuGo's score estimator, and
> >> trust it. Some of them will be wrong, I know, but most of them will be
> >> correct, so it is a sacrifice I am willing to make. The only difficulty
> >> I can foresee is knowing how the game ended -- GnuGo will give me an
> >> estimate, but it cannot know if White or Black resigned at some point,
> >> for example. My question, really, is if anyone knows of a program or
> >> script that will tell whether a game is completed -- that is to say, it
> >> is at a stage where only dame points are left, all territory is decided,
> >> etc. This way, I can run a game through both GnuGo and that program, and
> >> if the game is not finished, I will take GnuGo's estimation as a
> >> +Resign, and if it is, I will trust GnuGo's score estimate.
> >
> >You could ask GNU Go to generate a move, and then find out (e.g. via
> >top_moves) about it's value. If the move was pass, or it's value less
> >than 0.5 or so, the game should have ended.
> >
> >GNU Go' estimate will give the correct result for finished games with a
> >pretty good probability. For unfinished games, it will very often be
> > wrong.
> >
> >Arend
> >
> >_______________________________________________
> >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/