David J Wu

And I uploaded a pre-trained model here, if you want to try out the neural net without training your own. https://github.com/lightvector/GoNN/releases/tag/v0.1

MCTS implemented! Many new kinds of experiments to try. And some new weights: https://github.com/lightvector/GoNN/releases/tag/v0.2

New repo with working AlphaZero-like self play training! Further research will probably take place there, rather than here. :) https://github.com/lightvector/KataGo

Thanks! I will make this correction.

You want to download a GUI program, like KaTrain, q5Go, Lizzie, or other possible choices. KaTrain will handle a lot of the details for you, and might be one of...

Thanks for the result! I haven't done detailed strength testing, but yes your results are consistent with some more recent neural net training. Fixup underperforms batchnorm a bit. So I...

Here's an example plot I dug up from some old short training experiments, comparing fixup (blue) and batch norm (orange). No Elo testing, just visualizations of the policy loss during...

I don't know of a good method to deal with it. Even RL-trained nets can have a lot of difficulty with large dragons - KataGo can misevaluate them too. They...

Thanks. Is this in response to the recent ratings of the different networks, in particular the network after it seeming to be much worse? I didn't realize at first from...

Because you asked it to analyze turn 0 of that game record. `"analyzeTurns":[0]`. Turn 0 of that game record is an empty board because Q16 isn't played yet.