ReinforcementLearning.jl

ReinforcementLearning.jl copied to clipboard

The number of 'unmaintained' policies and significant subject-matter knowledge required to implement and verify correctness of many of the policies is making further progress on the v0.11 refactor difficult. I propose the following:

1. Add a new package (I know, not ideal) to this repository, ReinforcmentLearningGarden.jl, which will contain well-maintained policies / explorers.
2. Move policies one-by-one from RLZoo to RLGarden.jl
3. At some point deprecate RLZoo once everything has either been discarded or migrated

An alternative would be to drop everything which isn't of a certain quality / functionality level from RLZoo and add things back as we go, but the disadvantage here is that we lose a clear overview of what RLZoo contained / what could still be migrated.

@HenriDeh thoughts?

I would expect such a transfer to last for a very long time. During that time we would have two packages that do the same thing. This can be very confusing.

#614 contains a list of policies that need to be refactored. Maybe, in the meantime, we could mark them as broken (and invite the user to contribute ?) with a @warn instead of plain dropping them from the repo.

If making a v0.11 release without refactoring all policies is what you want to do with this, I think we can do it without a new package. The best we can do is be upfront to the user about the state of each algorithm (with the @warn and I think also in the docs and the Readme.md).

Sounds like a good idea.