This release includes several updates and improvements to the MARLlib. Below is a list of the major changes:
Added
- [API based usage](https://github.com/Replicable-MARL/MARLlib#getting-started): Since version 1.0.0, MARLlib has adopted a new API to facilitate training and enable better customization. For detailed usage, please refer to the [README](https://github.com/Replicable-MARL/MARLlib#getting-started).
- [Examples](https://github.com/Replicable-MARL/MARLlib/tree/master/examples): Examples are now provided to guide MARLlib users.
- [Render](https://github.com/Replicable-MARL/MARLlib/blob/master/examples/load_and_render_model.py): Pretrained models can be loaded and tasks can be rendered based on them.
- [Awesome list](https://marllib.readthedocs.io/en/latest/resources/awesome.html): A survey of recent papers has been included to showcase recent progress on MARL (up to early 2023).
Changed
- [MLP module](https://github.com/Replicable-MARL/MARLlib/tree/master/marllib/marl/models/zoo/mlp): Multi-Layer Perceptron is now available for building an agent architecture in this version, in addition to the previously available GRU.
- [Standardized encoder](https://github.com/Replicable-MARL/MARLlib/tree/master/marllib/marl/models/zoo/encoder): All different model classes now use the same encoder class.
- Fine-tuned hyperparameters: A [test folder](https://github.com/Replicable-MARL/MARLlib/tree/master/marllib/marl/algos/hyperparams/test) has been added for testing the algorithm and some environment folders have been deleted as their impact on parameters is not significant.
- Results: Some experiment data has been removed after checking.
- Other minor improvements.
Removed
- Console-based usage: Console-based usage has been completely deprecated as the new API is the only maintained way of using MARLlib in future development.
We encourage users to update to this version to take advantage of the new features and improvements. If you have any questions or issues, please don't hesitate to open an issue.