KataGo is a bot trained using a AlphaZero-like pipeline with a variety of experimental modifications and enhancements, but still learning only from playing itself starting from random play. It is currently using a 30 block neural net.

(2020-02-26) Trying a 30 block net.

Code: https://github.com/lightvector/KataGo

Paper: https://arxiv.org/abs/1902.10565

Blog: https://blog.janestreet.com/deep-learning-the-hardest-go-problem-in-the-world/ https://blog.janestreet.com/accelerating-self-play-learning-in-go/

Enjoy!