It is the first main version support for JP-like (Japenese-like) rule. We introduce JP rule game in self-play training so the weights could understand JP rule well. Note Sayuri can not understand Seki. She may be wrong in some positions. Please add --scoring-rule territory option and load the last weights to enable it. Enjoy!
Changes in v0.9.0
- Fix bugs for Python engine.
- Improve Python engine.
- Improve self-play performance.
- Add support for GTP command
sayuri-setoption. - No long support for
--relative-rankoption.
Run, 2024-12-5
We add the Policy Surprise Sampler the current Run, 2024-12-5. With this new technique, the last 15 block weight is already stronger than ELF OpenGo V0, a strong 20 block weight. This run also includes checkpoint file for Python engine. Please see the model page.
Full Changelog: v0.8.1...v0.9.0