[computer-go] The dominance of search (Suzie v. GnuGo)

[Don Dailey]

To take a normal evaluation function and convert it to a 
"probability of winning" function is probably difficult to 
do well.   You might have to map some sort of curve where
a few stones ahead represent a near win.

A simple approximation: - call the evaluation function - if
it is less than zero, consider it one monte carlo run where
white wins.    If it's greater than zero, consider it a 
monte carlo run where black wins.   

It's probably better to use the curve - but my sense of it
is that you need to map scores fairly accurately to odds
of winning - to truly simulate a monte carlo run.

How did you do this before?  I think almost any reasonable
idea is worth 2 or 3 tries - the devil is in the details.

- Don


 

On Sat, 2007-04-07 at 19:52 +0200, Chrilly wrote:
> >
> > I have this idea that perhaps a good evaluation function could
> > replace the play-out portion of the UCT programs.  The evaluation
> > function would return a value between 0 and 1 and would be an
> > estimate of the odds of winning.
> >
> I have tried this with an older and much weaker version of Suzie. It
> played 
> positionally better than the Alpha-Beta version, but the rate of very 
> strange moves also increased. UCT greates a more unbalanced tree than 
> Alpha-Beta and the programm has therefore even more chances to
> "cheat". For 
> the same reason extensions do not work so far in Suzie.
> 
> But I tried not with 0-1 but used the full eval. Maybe I should give
> it a 
> second try. But as I work now 45 hours/week on Computer-Tomography
> (which is 
> also quite interesting) and comute each weekend between Germany and
> Austria 
> its difficult to do.
> 
> Chrilly
>