As a long-time lurker I guess this is as good a first post as any.
I encountered this position some time ago (back in June I think), while pitching one of KataGo's distributed networks against Leela Zero. KataGo is Black and Leela is White. I didn't post this earlier because I was interested to see, if the training would rectify it on its own. But as it is, even the current networks enter into this variation (despite high playouts) and do not realize that the ladder doesn't work until it's too late.
Interestingly enough, the final network of the pre-distributed run (b40c256x2-s509) has no problems at all with this ladder. Is it possible that there is a risk of introducing new blind spots as the training goes on?
I think this is the move in question:
The image above shows the analysis of the currently highest-rated 40b network (b40c256-s1049) and it suggests a game-losing blunder after more than 500k playouts.
I remember lightvector saying (on this forum I think) that he could insert these kind of positions into the training, so I hope this is useful.
The sgf file contains only one variation (the original I encountered), but some networks might propose a different order of moves.
| Attachments: |
KataGo Blindspot.sgf [825 Bytes]
Downloaded 317 times |