Deciding training configuration

[drkostas]

Use previous actions as state dimensions

I know we haven't set up a full training example yet but the MalmoMazeEnv class has the option to record the action history and I was thinking that maybe we could use the last (or last 2) actions as state dimensions. Can be useful after "turn the camera" actions.

[drkostas]

Full Malmo Documentation

I just found the full documentation of Malmo that contains a lot of useful information: https://microsoft.github.io/malmo/0.30.0/Python_Examples/Tutorial.pdf

[drkostas]

https://microsoft.github.io/malmo/0.30.0/Schemas/Mission.html

[drkostas]

XML examples: https://github.com/microsoft/malmo/tree/master/sample_missions

[drkostas]

Minecraft blocks (to be used in the XML): https://minecraft-ids.grahamedgecombe.com/

[drkostas]

Hierarchical Training using ray example: https://github.com/ray-project/ray/blob/master/rllib/examples/hierarchical_training.py

[drkostas]

Ray Documentation: https://docs.ray.io/en/latest/rllib/core-concepts.html