alt113 / nl-fun

N-Layered FeUdal Networks based on FeUdal Networks adapted to suit PySC2 observations

NL-FuN

• Recreate baselines produced by O. Vinyals et al (2017) in StarCraft II: A New Challenge for Reinforcement Learning
• Modify FeUdal Networks (FUN) by A. S. Vezhnevets et al (2017) to suit the PySC2 observations.
• Generalize FUN to additional layers.

Use

I've added .bat files with examples of how to run the train.py file (need to change its name). The .bat files produce a shell command for each worker specified. Add --linux to produce .sh files instead. Use --python_v [python_cmd] to specify what command to run python with. For example: --python_v python3 if you have both python 2.x and python 3.x installed.

References:

Papers:

A3C: https://arxiv.org/pdf/1602.01783.pdf
PySC2 + Baselines: https://arxiv.org/pdf/1708.04782.pdf
FeUdal Networks: https://arxiv.org/pdf/1703.01161.pdf

Repositories:

Working on this project would not be possible without being able to use the following projects as references:

PySC2:

https://github.com/deepmind/pysc2

A3C:

https://github.com/dennybritz/reinforcement-learning/tree/master/PolicyGradient
https://github.com/xhujoy/pysc2-agents
https://github.com/awjuliani/DeepRL-Agents/blob/master/A3C-Doom.ipynb
https://github.com/chris-chris/pysc2-examples/tree/master/a2c

A3C + FullyConv:

https://github.com/H-Park/starcraft2ai/tree/master/A3C
https://github.com/pekaalto/sc2aibot

A3C + Distributed Tensorflow

https://github.com/openai/universe-starter-agent/blob/master/a3c.py

Feudal Networks:

https://github.com/dmakian/feudal_networks

Packages:

PySC2 == 1.2
tensorflow-gpu == 1.9 Python 3.x