Package Dependency from unity-ml is outdated

This is mostly from an upstream issue https://github.com/Unity-Technologies/ml-agents/issues/1646, but I believe it’s also worth being tracked here as well.

unity-ml requires a very specific, outdated version of tensorflow, numpy, etc. (and python is forced to >=3.6 <3.7, which shall be just 3.6+: #19). This would limit users to use recent version of softwares when developing a RL agent on their own, despite there is no reason to get tied to such old versions.

The most standard way: Have unity-ml fix https://github.com/Unity-Technologies/ml-agents/issues/1646 (e.g. split the monolithic one into env-related part and learning-related part). But, what else workaround we can think about?

One workaround is simply removing unity-ml from its requirements (setup.py), and let users manually install this with pip install unity-ml --no-deps, etc. Or we can consider mentioning in the README as well.

[awjuliani]

Hi @wookayin

Thank you for bringing this to our attention. I completely agree that these package requirements are overly restrictive. The current issue is that we are treating both the mlagents.learn and mlagents.envs packages as one, and using one set of requirements. In fact for Obstacle Tower only mlagents.envs is needed, and that package has very few dependencies. I will talk with my team about the best way to approach this in the short term for the competition.

[awjuliani]

@iamchathu Any Python version in the 3.5 or 3.6 range should work.