[Computer-go] global-pro-move-matching (was Recursive Neural Networks)

Darren Cook darren at dcook.org
Wed Jan 23 15:53:46 PST 2013


> E.g., they report that the top-ranked move from their neural net has a 10%
> chance of matching a pro move on a 19x19 board....
> And if you use a large-scale pattern approach (e.g., Remi's work), then you
> can predict over 40% and maybe up to 50% of pro moves.

This is a poor measure of strength though (which is why most go
programmers concentrate on win-rates against Gnugo, or KGS ratings, etc.)

Exhibit 1:
  I chose 3,4 in top-right corner for my first move; the pro chose 4,4.
Mark it down as just as bad as choosing the 1,1 point.

Exhibit 2:
  I chose the hane-tsugi sente sequence at A10. The pro first played the
monkey jump sente sequence in the top-right then played A10. Mark my
choice as just as bad as poking my own eye out.

Darren




More information about the Computer-go mailing list