Fast Reinforcement Learning Versions Save

The library as it is allows for easy training for DDPG based models and DQN based models. You can also save them, and reload them.

You can use the different interpreter objects for graphing rewards, comparing rewards with other models, viewing episodes at different periods of the agent's training, etc.

Notes: Currently, the next obstacle is memory efficiency. We will be adding more models, but will also be addressing memory issues possibly by off loading to storage.

All gifs are added. There may be a few extra added in later versions, however all of the base env runs are there.

Still does not contain Gifs. This is primarily a test of the azure pipeline publishing packages for us. Once a PR is pushed to master, the new version will automatically be updated in PyPI.

Next Release will have Gifs, then soon after redone readme

Some key take aways with this release:

• 75% of the code has been trashed and replaced. Code is now organized more similarly with the fastai gan.py file as well as the computer vision models.
  • Uses fasti tabular model for embeddable and linear state / general inputs
• jupyter notebooks with executable code displaying graphs.
• More interpreter improvements

Why the changes?

• Main motivation is making the "trained" models as separate from the rest of the fastrl API. This means that the models are not longer having data / learner / Action / State objects embedded inside them. This should be portability easier. The main code base that is heavily integrated with fastrl will be part of the training, but you could / will be able to run the models without fastrl once trained (in theory, we'll see)

Right now some basic model configurations are complete with unit tests. Moving forward, we will be checking the model performance that is expected on a set of environments.