Counting 1,690 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Author
Contributors
Last Commit
Aug. 18, 2017
Created
Jul. 25, 2017

PySC2 - StarCraft II Learning Environment

PySC2 is DeepMind's Python component of the StarCraft II Learning Environment (SC2LE). It exposes Blizzard Entertainment's StarCraft II Machine Learning API as a Python RL Environment. This is a collaboration between DeepMind and Blizzard to develop StarCraft II into a rich environment for RL research. PySC2 provides an interface for RL agents to interact with StarCraft 2, getting observations and sending actions.

We have published an accompanying blogpost and paper, which outlines our motivation for using StarCraft II for DeepRL research, and some initial research results using the environment.

About

Disclaimer: This is not an official Google product.

If you use the StarCraft II Machine Learning API and/or PySC2 in your research, please cite the StarCraft II Paper

You can reach us at [email protected].

Quick Start Guide

Get PySC2

PyPI

The easiest way to get PySC2 is to use pip:

$ pip install pysc2

That will install the pysc2 package along with all the required dependencies. If you're running on an older system you may need to install libsdl libraries for the pygame dependency.

Pip will install a few of the binaries to your bin directory. pysc2_play can be used as a shortcut to python -m pysc2.bin.play.

Git

Alternatively you can install PySC2 with git. First clone the PySC2 repo, then install the dependencies and pysc2 package:

$ git clone https://github.com/deepmind/pysc2.git
$ pip install pysc2/

Get StarCraft II

PySC2 depends on the full StarCraft II game and only works with versions that include the API, which is 3.16.1 and above.

Linux

Follow Blizzard's documentation to get the linux version. By default, PySC2 expects the game to live in ~/StarCraftII/. You can override this path by setting the SC2PATH environment variable or creating your own run_config.

Windows/MacOS

Install of the game as normal from Battle.net. Even the Starter Edition will work. If you used the default install location PySC2 should find the latest binary. If you changed the install location, you'll need to set the SC2PATH environment variable with the correct location.

PySC2 should work on MacOS and Windows systems running Python 2.7+ or 3.4+, but has only been thoroughly tested on Linux. We welcome suggestions and patches for better compatibility with other systems.

Get the maps

PySC2 has many maps pre-configured, but they need to be downloaded into the SC2 Maps directory before they can be played.

Download the ladder maps and the mini games and extract them to your StarcraftII/Maps/ directory.

Run an agent

You can run an agent to test the environment. The UI shows you the actions of the agent and is helpful for debugging and visualization purposes.

$ python -m pysc2.bin.agent --map Simple64

It runs a random agent by default, but you can specify others if you'd like, including your own.

$ python -m pysc2.bin.agent --map CollectMineralShards --agent pysc2.agents.scripted_agent.CollectMineralShards

To specify the agent's race, the opponent's difficulty, and more, you can pass additional flags. Run with --help to see what you can change.

Play the game as a human

There is a human agent interface which is mainly used for debugging, but it can also be used to play the game. The UI is fairly simple and incomplete, but it's enough to understand the basics of the game. Also, it runs on Linux.

$ python -m pysc2.bin.play --map Simple64

In the UI, hit ? for a list of the hotkeys. The most basic ones are: F4 to quit, F5 to restart, F9 to save a replay, and Pgup/Pgdn to control the speed of the game. Otherwise use the mouse for selection and keyboard for commands listed on the left.

The left side is a basic rendering (which will likely be replaced by a proper rendering some day). The right side is the feature layers that the agent receives, with some coloring to make it more useful to us.

Watch a replay

Running the random agent and playing as a human save a replay by default. You can watch that replay by running:

$ python -m pysc2.bin.play --replay <path-to-replay>

This works for any replay as long as the map can be found by the game.

The same controls work as for playing the game, so F4 to exit, pgup/pgdn to control the speed, etc.

List the maps

Maps need to be configured before they're known to the environment. You can see the list of known maps by running:

$ python -m pysc2.bin.map_list

Environment Details

A full description of the specifics of how the environment is configured, the observations and action spaces work is available here.

Mini-game maps

The mini-game map files referenced in the paper are stored under pysc2/maps/ but must be installed in $SC2PATH/Maps. Make sure to follow the download instructions above.

Maps are configured in the Python files in pysc2/maps/. The configs can set player and time limits, whether to use the game outcome or curriculum score, and a handful of other things. For more information about the maps, and how to configure your own, take a look here.

Replays

A replay lets you review what happened during a game. You can see the actions and observations that each player made as they played.

Blizzard is releasing a large number of anonymized 1v1 replays played on the ladder. You can find instructions for how to get the replay files on their site. You can also review your own replays.

Replays can be played back to get the observations and actions made during that game. The observations are rendered at the resolution you request, so may differ from what the human actually saw. Similarly the actions specify a point, which could reflect a different pixel on the human's screen, so may not have an exact match in our observations, though they should be fairly similar.

Replays are version dependent, so a 3.15 replay will fail in a 3.16 binary.

You can visualize the replays with the full game, or with pysc2.bin.play. Alternatively you can run pysc2.bin.replay_actions to process many replays in parallel.

Latest Releases
v1.1
 Aug. 15 2017
Initial release
 Aug. 9 2017