Pysc1-gym

Installing and managing Starcraft I learning environment can be tedious and full of errors. Here, I present to you packaged way of running a nice Starcraft I gym simulation in your laptop or some clusters.

I was using Windows 10 while making this work, so the instructions are made in reference to the Windows system.

“OpenAI Gym” like code for Starcraft I

After the Environment is set, you can access the features of Starcraft I just like in OpenAI Gym:

env = sc.SingleBattleEnv(args.ip, args.port)
env.seed(777)
agent = RandomAgent(env.action_space)

episodes = 0
while episodes < 100:
    obs = env._reset()
    done = False
    while not done:
        action = agent.act()
        obs, reward, done, info = env._step(action)
    episodes += 1

env.close()

prerequisites

• OpenAI Gym
• Starcraft 1.1.16
• BWAPI 4.1.2
• TorchCraft 1.0.2
• TorchCraft-py
• Gym-TorchCraft

Because of the dependencies here, Python 2.7 is used. In the future, support for Python 3.6+ would be added.

• Python 2.7 I recommend using conda create -n "pysc1-gym" python=2.7 to manage separate packages.

Get the TorchCraft 1.0-2 from here

check to prerequisites again. For start, git clone https://github.com/openai/gym and cd gym and pip install -e . for minimal install of gym.

Installation (Windows 10 with Anaconda)

The installation process is long, but it should be lot shorter than doing this from scratch. So relax and slowly follow the steps. Take a deep breath…

1. activate pysc1-gym to activate the environment.

2. Check your Starcraft Installation. The directory is called $STARCRAFT.

3. run pysc1-gym/bwapi/BWAPI_412_Setup.exe and install on $BWAPI directory and run $BWAPI/Chaoslauncher/Chaoslauncher.exe that is inside the BWAPI installation location.

4. From Settings in Chaoslauncher, choose the path to the Starcraft installation directory $STARCRAFT.

   4.1. Full Installation

   4.2. The Abyss (ICCUP) and Games Podolsk are optional

5. copy files in pysc1-gym/torchcraft/*.dll to $STARCRAFT directory.

6. copy files in pysc1-gym/config/*.ini to $STARCRAFT/bwapi-data directory

7. run pysc1-gym/install.bat (while still inside the pysc1-gym environment). This will install custom dependencies torchcraft-py and gym-starcraft. (Check if no existing installations are there in the pysc1-gym env.) This will take some time installing the dependencies. wait patiently… have a coffee.

8. copy contents of pysc1-gym/maps/ to $STARCRAFT/maps/.

9. run Chaoslauncher with BWAPI 4.1.2 injector [RELEASE] and W-mode 1.02 checked. The Starcraft window will open in waiting mode.

   Don’t panic if the window is not responsive. It’s supposed to be that way!

10. cd pysc1-gym/examples (while still inside the pysc1-gym environment) and run python random_agent.py --ip=localhost --port=11111

Custom Maps for Reinforcement Learning

Custom maps I created for diverse RL environments.

BroodWar/Micros

• bc12v12_c_far.scx 12 Battlecruisers against each other

• m24v24_c_far_mod.scx 24 Marines against each other

• zt12v12_j_far.scm 11 Zealots and a High Templar against each other

TODO

• support python 3.6+
• more easy installation.