RE: computer-go: Applying Moore's Law to Computer Go

[Darren Cook <darren@xxxxxxxxxxxxxx>]

>* How I wish my program can evaluate 1 node per second. :) My feeling is
>that if my program can evaluate 10^5 nodes in a couple of minutes, it's
>hopeful to get IGS *2k, and thus win a tournament. My problem might be

Here's how you do it:
   Search to the end of the game
   Have a very simple evaluation function: count territory

So spread your 10^5 nodes, as 500 x 200-ply searches. That implies that
almost every node will have a branching factor of 1.

I'm serious: an accurate middle game evaluation function for go is very,
very, very difficult. Intelligent move ordering is merely "very difficult".
So it makes sense to focus efforts on the latter. [1]

The deeper you are in the tree the safer the moves you play - you always
defend a possibly weak group rather than attack an enemy to avoid
life/death battles.

This is also how humans play. When you count in the middle game what you
actually do is resolve life/death quickly then imagine playing all the
obvious end game moves (the ones that avoid complex variations), and then
just count territory.

Darren

[1]: Chess is probably the reverse: evaluation is easier than move
ordering, so you use a full-width search. FWIW, I think shogi is more like
go, because position is more important than material, so the evaluation
function is harder.