版本发布日志¶

XuanCe (1.4.1)¶

Released on 2026-02-25 - GitHub - PyPI

Updated: Upgraded some Box2D scenarios to the latest versions.
Improved: Further optimized the Benchmark pipeline/workflow.
Optimized: Improved and tested image-input support for MARL algorithms.
Updated: Refined the config system to simplify the parameter setup process and improve readability.
Refactored: Moved Runner to a backend-agnostic (backend-independent) layer.
Renamed: dl_toolbox → dl_backend.

Full Changelog: 1.4.0...1.4.1

Released on 2026-01-12 - GitHub - PyPI

Fixed: Resolved installation issues by removing the hard dependency on mpi4py, significantly simplifying the installation process.
Changed: Moved set_seed functionality from the Runner module to the Agent module for clearer responsibility separation.
Improved: Optimized the design of test_envs to reduce unnecessary resource consumption during evaluation.
Improved: Refactored the Runner module to better manage Agents, environments, and experiment lifecycles, enabling clearer and more consistent workflows for training, testing, and benchmarking.
Improved: Further standardized the benchmark pipeline to make it easier for users to quickly run their own benchmarks and obtain more comprehensive benchmark results (work in progress).
Refactored: Updated the Agent class to support metadata persistence and revised its initialization interface.
Renamed: Renamed the Runner parameter method to algo for clearer semantic meaning.
Fixed: Various other bug fixes and stability improvements.

Released on 2025-12-31 - GitHub - PyPI

Docs: Update documentation (thanks to @S444H @1otuses @jiaweiLu04 @YeFanRepo @Gaoshen-worker @GUOZI-fruit @zhanshuxie).
Callbacks: Improved callbacks for MARL and other AI toolboxes.
Models: Added support for CNN-based representations in MARL algorithms.
Environments: Added Atari environments for MARL.
Benchmarks: Added and standardized the Benchmark section for XuanCe.
Algorithms: Added support for Independent TD3 (ITD3) in MARL.

Full Changelog: 1.3.2...1.3.3

Released on 2025-10-02 - GitHub - PyPI

Synchronize updates for the TensorFlow and MindSpore frameworks by @wenzhangliu
Add support for running on Colab by @wenzhangliu
Fix the learning rate decay step issue (no decay by default) by @wenzhangliu
Docs(curl_agent.md, drq_agent.md, spr_agent.md): add docs. by @wenboli-ai in #155
Configurable parameter sharing for MADDPG adversarial environment (add new examples) by @josh1147 in #156
Add custom MARL policy implementation example by @josh1147 in #157
Update drqn_agent.py by @1otuses in #158
Fix other bugs.

Full Changelog: 1.3.1...1.3.2

Released on 2025-07-02 - GitHub - PyPI

Comprehensive testing completed for Python 3.8, 3.9, 3.10, 3.11, and 3.12. @wenzhangliu
Dropped official support for Python 3.6 and 3.7. @wenzhangliu
Adjusted version requirements for commonly used packages: numpy, torch, gymnasium, etc. @wenzhangliu
Update MARL communication algorithm by @TangY1fan in #150

Full Changelog: 1.3.0...1.3.1

Released on 2025-06-17 - GitHub - PyPI

callback: support callback functions by @wenzhangliu, allowing users to insert new code blocks
fix: test results miscalculation in marl settings by @yuzh2001 in #111
docs: fix recent section links by @guspan-tanadi in #114
add dreamerv3 implementation in example by @LKP1111 in #125
Add TD3_BC algorithm by @XiangDuojie in #129
add ic3net by @TangY1fan in #130
harmonydream by @LKP1111 in #131
fix ic3net test by @TangY1fan in #132
add dreamerv2 by @LKP1111 in #133
fix bugs for ic3net by @TangY1fan in #134
Basic_ViT layers by @wenboli-ai in #139
adapt some vit-related file by @wenboli-ai in #140
add commnet and update ic3net by @TangY1fan in #142
update a bit misspell by @guspan-tanadi in #143
add Contrastive Reinforcement Learning by @wenboli-ai in #144
add an algos of CRL (Data-Efficient Reinforcement Learning with Self-Predictive Representations) by @wenboli-ai in #145
add TarMAC by @TangY1fan in #146
add new algorithm of dgn by @TangY1fan in #147
feat(drq_agent, drq_learner, drq_atari): add an algorithm(DrQ). by @wenboli-ai in #148
update CommNet and IC3Net by @TangY1fan in #149
fix Atari environment wrapper (add XuanCeAtariEnvWrapper) and fix hidden bugs in the original AtariEnvWrapper by @wenzhangliu
remove gym by @wenzhangliu, now using the latest version of gymnasium (v1.1.1)
fix and update the mean field reinforcement learning algorithms: MFQ and MFAC by @wenzhangliu

Full Changelog: 1.2.6...1.3.0

Released on 2025-02-08 - GitHub - PyPI

The Optuna tool was integrated to support automatic hyperparameter tuning by @wenzhangliu.
Add NPG agent and NPG learner by @XiangDuojie in #102
Optimized COMA for tasks on SMAC by @wenzhangliu.
Added the get_joint_input method to the LearnerMAS class to fix data reading failures when the number of agents is 1 by @wenzhangliu.
Other bug fixes.

Full Changelog: 1.2.5...1.2.6

Released on 2025-01-05 - GitHub - PyPI

Updated and optimized the API documentation.
Updated the Atari environment.
Correct the env argument of Agent as envs: Union[DummyVecEnv, SubprocVecEnv].
Updated the RNN support in MASAC.
Some gym environments are modified to be accessed via the gymnasium interface.
Adjusted the dependency package versions during the XuanCe installation process and added support for installation for specific environments.
Other bug fixes.

Full Changelog: 1.0.0...1.2.5