Description
Tianshou is nearing its next major release and there are several endeavors in place to popularize it among new users in industry and research. The upcoming talk at PyData Berlin (the only talk on RL there) is one example.
A more complete and well-formulated documentation plays a crucial role in extending the tianshou community. This issue is an invitation to everyone to contribute to the documentation. Current issues and missing pieces include
- Duplication in content between tutorials in .rst and notebook tutorials. Notebook tutorials should be preferred as delivery mechanism.
- Information on high-level API and examples in the docs (currently just in readme)
- Notebook tutorial on high-level and the IDE usage based on naming conventions
- Docs on how to run and interpret a training-eval protocol (needs some code for that to be committed first)
- Docs on how to base your own policy class using tianshou's abstractions, how to train it, how to include it in high-level interfaces
- Basic ideas on when and why researchers should prefer tianshou to non-modularized frameworks (like cleanRL)
- Grammar, structure and phrasing in many places over the docs is not too good and should be improved
- Docs on overall tianshou development goals
- Docs on what may be broken (MARL, LSTM things) and what works well
- Updated comparison to other RL frameworks, when and how to use tianshou vs other things
- Highlighting some design principles with examples (modularity, prototype-based typing where appropriate, declarative high-level APIs, statistically significant performance evaluations)
- And many more things
If you have time to spare, it is usually a task that requires relatively little focus to go over some place in the docs and improve it a bit. Such improvements will be reviewed and merged quickly and when accumulated will provide a massive value to the project!
Now the docs delivery mechanism involves automatic QA on notebooks (execution, spelling, types, auto-formatting) and adding to the docs is easier and safer than ever!
I'm pinging all who actively contributed to the code or discussion in the last months - if you have some time to spare, consider helping out here :). You can also ping more people in the comments if you think they'd like to contribute.
This issue is just an overview, ping me to extend the above list and feel free to create new issues related to documentation to keep track of tasks and progress.
@Trinkle23897 @nuance1979 @spacegoing @carlocagnetta @dantp-ai @opcode81 @bordeauxred @arnaujc91 @maxhuettenrauch