We strongly encourage researchers to provide us feedback and contributing to MushroomRL. You can contribute in the following ways:
- providing bug reports;
- implementing new algorithms, approximators, environments, and so on;
- proposing improvements to the library, or documentation.
Please use GitHub Issues and use the "bug" tag to label it. It is desirable if you can provide a minimal Python script where the bug occurs. If the bug is confirmed, you can also provide a pull request to fix it, or wait for the maintainers to resolve the issue.
Although any contribution is welcome, we will only accept high-quality code that will reflect the key ideas of modularity and flexibility of MushroomRL. Please keep this in mind before submitting a pull request.
Every algorithm in MushroomRL must extend the Agent interface. If the algorithm belongs to a class of algorithms already
implemented in MushroomRL, then we expect the author to extend the appropriate interface. If new policies or new parameter types
are needed, please extend the corresponding base classes. Using other MushroomRL modules (e.g. Regressor, Features) is
highly encouraged. Together with the algorithm, you must provide an example and a test case. Every algorithm matching our
standard of code and scientific quality and providing the example and the test, will be merged in MushroomRL. It is encouraged
to include the citation to the paper describing the algorithm in the docstring of the class.
MushroomRL supports the use of third-party function approximators (e.g. Scikit-learn, Keras),
or customized ones, given that they implement the fit and predict methods.
The environment template of MushroomRL follows the style of the one introduced in OpenAI Gym.
Customized environment can be added provided that this exact style is used, and extending
the Environment interface. Customized functions for visualization, i.e., render, and
others can be easily added.
MushroomRL provides several examples showing the expected instructions sequence that an experiment should have. This structure is not mandatory, but recommended. Although all the examples share some similarities, there are some differences depending on the algorithm and environments at hand, and also on the specific needs of the user. It is desirable to propose an example whenever a new algorithm or environment is proposed in a pull request. The choice of the type of experiment is up to the contributor, but it should be meaningful to show the performance and properties of the proposed algorithm/environment.
Instructions for adding a new benchmark are provided in the MushroomRL Benchmarking Suite.
If you find some lack of features, please open a GitHub issue and use the "enhancement" tag to label it. However, before opening a pull request, we suggest to discuss it with the maintainers of the library, that will be glad of providing suggestions and help in case the proposed improvement is particularly interesting.