Train and benchmark several RL algorithms

[ghost]

🚀 Feature Request

Part of #42. Depends on #44. Once an environment is set up, it will be easy to train several of the RL algorithms provided by pytorch. All of these algorithms should be bench marked and a team discussion take place on which one to use for production training. The computation library for performing these tasks will be caffe2 as it is easy to deploy on production cloud services. The focus is on 2019: creating a generic tool for this is not essential, but it will be very beneficial for future years, and the task of #51.

[Ivan-Z]

Are we for sure going for Keras over Pytorch?

[ghost]

Yes, keras-rl lets us run tensorflow which is never going anywhere, and is also pretty damn fast. So it gives us speed and security.

[ghost]

We are now going to be using pytorch for several reasons, after talking to industry experts on machine learning. The main reasons are:

1. Faster development speed. The number one reason highlighted by every expert, regardless of whether they preferred pytorch or not, is that pytorch has undoubtedly the fastest development cycle of any machine learning library.
2. Brighter future: pytorch is growing faster than tensorflow/keras and is expected to continue. It's development community is, at the moment, growing faster as well, and with it comes exceptional libraries and support.