All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
1.1.0 - 2024-04-XX
- Modern build system using
pyproject.tomlwith scikit-build-core - cibuildwheel GitHub Actions for cross-platform wheel building
- Trusted publishing to PyPI via OIDC
- Sphinx documentation with Furo theme
- GitHub Pages documentation deployment via GitHub Actions
- Python 3.13 and 3.14 support
- Python version support updated to 3.10-3.14 (dropped 3.8, 3.9)
- pybind11 is now a build dependency instead of vendored
- Build backend changed from custom setup.py to scikit-build-core
- CI workflow now tests multiple Python versions (3.10-3.13)
- Vendored pybind11 from ThirdParty/ directory
- Complex CMakeBuild class from setup.py (now a shim)
1.0.5 - 2024-01-XX
- Migrated from gym to gymnasium for modern compatibility
- Fixed CI workflow for compatibility with modern GitHub Actions runners
1.0.4 - 2023-XX-XX
- Pretrained VLOG models (CQL and BC variants)
- Offline dataset for training available on GitHub Releases
- Various bug fixes and performance improvements
1.0.0 - 2022-04-XX
- Initial release
- Complete Japanese Riichi Mahjong rules implementation
- Multi-agent environment (
MahjongEnv) - Single-agent environment (
SingleAgentMahjongEnv) - Observation encoding with executor (93 channels) and oracle (18 channels) views
- Paipu replay system for Tenhou.net game records
- High-performance C++ backend with pybind11 Python bindings
- Pretrained opponent models for single-agent mode
- Introduced in ICLR 2022 paper: "Variational Oracle Guiding for Reinforcement Learning"