Learning about, deconstructing, and engineering new reinforcement learning problems in OpenAI gym.
The OpenAI gym has a nice and fairly extensive API for instantiating and training machine learning problems. However, it does not appear to reflect best practice from a controls and dnyamical systems perspective. the objective behind this repository is to both understand the API associated to the OpenAI gym, as well as to extend and augment it with more controls relevant implementations. This effort mirrors work of our collaborators in a control + learning library for python lyapy.
Hopefully we can have the first reasonable version done by March 2020. Fingers crossed. The reinforcement learning and the control + learning work is also being mirrored in Matlab by us. The Matlab version is currently private but will be released shortly after or around the python versions. Some effort will be made to align the implementations. So far it seems like Matlab has duplicated the overall class structure of the OpenAI gym implementation. If this is mostly the case, then one can use whichever programming environment is most comfortable. The controls + learning APIs will most likely differ somewhat.