Using StarCraft II for AI Research

Building a neural network that does things other than linear regression, MDP or identifying cats in images is apparently non-trivial. Not because of the complexities of the AI/DL/RL subject matter, but because working with raw datasets is atrociously boring. After the books and blogs with their examples, whats next to apply a neural network to? Games. 😎

For me, the following were candidates but each have been a challenge to get up and running either because of unsupported platforms, blocking bugs or just too complicated:

• Use the BizHawk game emulator to have a neural net play Super Mario Brothers (Runs on Windows only)
• Nvidia’s Isaac SDK and Sim (AFAICT requires access to an Nvidia card)
• FaceBook’s Habitat (Seemed too advanced with vision processing for what I wanted to do)
• Vincross Mind SDK with SIM (Seemed too vertical and specific to the Hexa platform)
• OpenAI Gym (This is a nice platform that I plan to spend more time with)

Then I found StarCraft II and AlphaStar. In 2016, DeepMind and Blizzard teamed up to make StarCraft 2 an AI research platform. See this Youtube video showing DeepMind’s AlphaStar play SC2 tournament pros.

Installation is a breeze:

1. Install Battle.net & StarCraft II (the free version)
2. pip install pysc2
3. python -m pysc2.bin.agent --map Simple64

And you get a running agent and a special dashboard for monitoring:

The ease and simplicity of getting this all working and running makes this a great option. Now for the fun stuff…

Related Links:

• DeepMind’s Python package for StarCraft II (pysc2)
• sc2ai Wiki Page
• AI defeats humans at StarCraft II