[computer-go] a diffcult position for UCT/MCGO

[Magnus Persson]

Quoting elife <elife2008 at gmail.com>:

> Hi all,
>  In the game attached, even after B played E2, a pure uct/mcgo program with
> 1.2M playouts will still evaluate the game as white win(the expected value
> is about -0.4). But in fact W is already lost. I guess it's because the
> program wrongly thinks it's very possible white will kill black's right down
> corner. Any comment?

The position is a 0.5 point win for white. White has a tesuji, to play 
A6 rather
than B6. Black can try play atari with A6 but then white gets sente 
with A5. The
shapes looks shaky for white and I instinctly rejected this variation but it
seems to work. I overlooked this myself, but Valkyria read this 12 ply deep in
5 minutes and just become more and more sure of a win for white which is not
normal for a losing position.

I think this is a good example when full board search easily finds odd looking
moves that strong humans (well at least as strong as me) might easily 
overlook.

-Magnus