Does not see a winning reply, even with 100k playouts

[dsmic]

The last move of w was not in the search tree, even with 100k playouts.

In the actual situation it was lost anyway, some moves before it was possible to win the game by killing the r4 group, if it has seen this move.

Reproduce: set move to 280, do 100000 playouts and output sgf

(at the moment I do not have a good understanding, why it does not get into the tree.)

[Francois van Niekerk]

Black is not seeing that the L2 stones are dead and incorrectly reading the very simple capturing race in the bottom right corner.

[dsmic]

This is extremly sensitive to progressive broadening. Turning it off or reducing parameter progressive_broadening_a to 2 makes oakfoam seeing the situation.

[Francois van Niekerk]

You probably mean progressive widening. This may be true, but generally progressive widening helps with strength. After the parameters are all re-tuned, from the ground up, this will have to be re-evaluated. Capturing races are often a problem for MCTS.

[dsmic]

Simple live and death at move 213 which is not seen with progressive widening catalin-NiceGo.sgf

[dsmic]

But you are right, without progressive widening it plays nearly random on 19x19 with 10000 ppm, I just tried:)

[dsmic]

Another quite simple situation of this problem. Kill in 2 moves not found:(

[dsmic]

I close this, as we have a lot of simelar and easier to understand samples