[computer-go] New version of CrazyStone

[Rémi Coulom]

Tom Cooper wrote:
> It didn't strike me at the time, but this is a remarkable 
> achievement.  Crazy stone seems to
> be competitive (to put it mildly) on a 7x7 board with a professional 
> go player!
> (Game records here 
> http://kgs.kiseido.com/en_US/gameArchives.jsp?user=CrazyStone 
> ).Congratulations
> to Remi.
Thanks
>
> I'd like to ask him and the other Monte-Carlo programmers if they have 
> any guesses about the
> relative computing effort required to produce a similar performance 
> using the same methods
> on an 8x8 board, a 9x9, a 13x13 or a 19x19.
A lot of the strength of Crazy Stone on 7x7 comes from its opening book. 
Even I manage to beat it on 7x7 from time to time when it does not play 
with an opening book. I have never tried 8x8. The method scales very 
badly to larger boards. I expect that on 9x9 no reasonable amount of CPU 
power would make Crazy Stone reach dan level.

The big problem is that the algorithm is a global tree search. My 
feeling is that even the best possible global tree-search algorithm 
cannot play well on 19x19, whatever excellent selectivity and evaluation 
function is used. It is necessary to invent new approaches.

Rémi