[computer-go] use for Monte Carlo on 19X19?

[Benjamin Teuber]

For scoring, you could use integral instead of boolean outcomes again, which
would solve some of the problems you descibed...
This might be a great idea for programs that must deal with Japanese
scoring, but for our reseach the best thing is just to stick to Chinese
rules :)

On 11/5/07, Don Dailey <drdailey at cox.net> wrote:
>
> It would be tricky indeed to implement Japanese scoring with Monte Carlo
> programs.
>
> Monte Carlo programs are  very strong and thus they are not naive about
> what the situation is.    But I view them as "relativistic players"  -
> whether a group is dead or alive depends on your point of view!   A
> group is dead if it needs to be dead to win.
>
> This makes it a difficult fit for scoring dead groups.   A monte carlo
> go program might even play a move that makes it possible to win your
> dead group back if it's not relevant to it's winning chances.     So
> dead group analysis must be done by external methods in my opinion.
>
> Also, Japanese scoring is problematic because it gives meaning to moves
> which should be considered meaningless.     Monte Carlo only cares about
> meaningful moves - moves which increase the winning chances.    Some of
> the moves that  Monte Carlo considers worthless are considered to be
> negative worth in  Japanese scoring.
>
> When a monte carlo program thinks the game is dead won or lost, you can
> stop playing - but you can't necessarily get a correct bean count of the
> position.
>
> So I don't think Japanese scoring is compatible with Monte Carlo methods
> - although with some cleverness I'm sure it's possible to impose it.
>
> It occurred to me that you could build your play-outs to use Japanese
> counting.    I'm sure this would weaken the program considerably
> however.    Japanese scoring would impose a lot of "noise" on the
> positions.
>
>
> - Don
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20071105/81d92008/attachment.htm