The branch game created from the previous failed game and then succeeded.
You can also try it from this branch game
6 Likes
Yay, eat eat eat! 
This one is probably going to be hard for micro to play correctly, even after more training, just because the position is so extreme and there are so many equivalent moves that a 40ish-visit search spread out between all the equivalent moves is not going to do very much.
But maybe someday. You can see from the chat that the value head still does see some danger ahead of time. I’m not entirely sure I want to add a lot of things quite this extreme to the training, which might give very large gradients due to the extreme swing and make the model’s training less stable. But maybe continued improved danger sense on large but not-whole-board groups will eventually get some amount of better whole-board-sized-group judgment for free via generalization, enough to where a larger visit search can handle it. Not sure!
Improvements still happening! I looked at many of the games back in March and April and at least the final losing move in many cases I checked are now much more likely to be handled correctly (although low playouts may still give chance for error).
4 Likes
6 Likes
1 Like
Unfortunately time out
another full game