wrong reading of a simple corner
Issue #18
resolved
bottom left is read wrong. I do not understand why this case is so special, usually it reads fine at the moment. I double checked, if I messed something up, but everything seems fine.
I even turned the poolrave on. It is even not finding the move in the search tree fast enough. Perhaps you have an idea, I think a lot of wrong reading in 13x13 was this type...
Comments (6)
-
repo owner
-
repo owner
- marked as enhancement
-
reporter
OK, the moves at b3 and a1 have the atari feature 3. I am not sure if this makes any sense, as it is given by
if (board->isLadderAfter(group,move) && !board->isProbableWorkingLadderAfter(group,move)) return 3;
atari feature 3 has no high value in your hand picked value and it was not much higer in my new trained gammas. At least the idea of ladder code should probably not cover this situation?!
-
reporter
Sorry, feature_ladder was turned off, so it is not the reason...
-
reporter
I was thinking it might be a problem of naked moves. We use them in playouts but we do not have a feature for the tree with this. Same for the more complicated 4 and 5 naked?!
-
reporter
- changed status to resolved
I close this, we have simelar examples in the regression and this one is read not too bad with the my current branch
- Log in to comment
The bottom left turns into a small semeai-like situation. I don't think this is a bug so much as a "known weakness." These sorts of situations are usually quite difficult to read out in simple playouts.