v4.0: Add the C++ converters used by AlphaStar

Add C++ environment converters which handle the conversion between the protos used by SC2 environment for actions, observations and specs to a more neural network friendly dictionary of Numpy arrays, with accompanying environment specs. Converters are built as a translation of relevant pieces from sc2_env.py from Python to C++ for speed.

New features include the following

• 'Raw' and 'visual' modes are supported explicitly.
• Actions are represented as simple dicts and have a delay member which specifies how many game loops to wait until the next observation is retrieved.
• The following new observations are available (parameterised by a config proto, see the converter documentation for more detail):
  • Action chosen (for supervised).
  • Own race requested
  • Unit counts and upgrades.
  • Camera, position and size (for raw, indicates where the camera is and when active also hides unit data outside the camera bounds).
  • Opponent race requested (and observed, once a unit has been seen)
  • Player stats vector, unit counts and upgrades.
  • MMR
• Minimap and screen features have an observation per plane rather than being stacked.
• Converters can be used both with Python and C++. The C++ code is wrapped using pybind11 facilitating Python use.
• Converters have been used in this recent work on StarCraftII Unplugged.
• Support building with Bazel as converters built on C++ requires compilation for any changes made to them.
  • With converters, PySC2 is a mixed Python and C++ repository.
  • Using setup.py will install the Python only components of the package.
  • Converters require either building with Bazel or can be pulled from pypi. Only Linux is supported at present.
• Drop Python 2 support, bump minimum version to Python 3.8+
  • Remove references to six, future, whichcraft.which.
  • Fix lint errors arising from py2-> py3 transformation.
• Fix fickle tests that fail because of obsolete code, proto dependencies, timeout or unavailable maps.
• Update PySC2 master documentation in line with these changes.
• Remove unnecessary patches that have been fixed in upstream repos.

v3.0: Radar is only defined if you use raw or feature units.

This is a fairly big release adding support for recent versions, many new observations, various new features, and bug fixes. In particular the raw actions are fairly well supported.

Changes:

• Add raw observations and actions
• Support the new actions, research, etc from the 4.7.1 and 4.10.1 balance patches.
• Many new observations:
  • alerts
  • hallucinations
  • cloaked
  • blips
  • buffs
  • effects
  • upgrades
  • active
  • pathable
  • buildable
  • rally
  • orders
  • cargo
  • radar rings
  • map name
  • races
  • full response observation protobuf
• Unit tests for some new observations.
• Support newer versions of the game
• Support the newest ladder maps.
• Support playing on battle.net maps.
• Support specifying the bot build
• Support training on random maps/races.
• Remove the long deprecated arguments to sc2_env.
• Force player arguments to be set and valid.
• Force player names to be unique to avoid an bug in the game where they control the same side.
• Connect to an existing instance
• Track and induce action delays to simulate latency
• Fix shrinking pygame window size.
• Fix deadlock in resizing the pygame window.
• Script to update your battle.net cache based on replays.
• Better error handling and error messages
• Better protocol logging
• Better benchmarking scripts and tooling
• Better version comparison scripts.
• Fixed upgrade names.
• Support multiple maps and races at env creation time.
• Added an example raw agent.
• Various fixes for new numpy versions
• Better support for slicing into NamedNumpyArray's.
• Support multiple actions per frame (discouraged)
• Better rendering:
  • Raw grids
  • selection, upgrades, orders, rally, placeholders, start locations, facing, radar rings
  • playable area
  • play with raw actions
  • Camera control during replays
• Add a basic ascii renderer
• Various bug small fixes.

Version 2.0.1

Fix a python 3 compatibility issue.

Version 2.0

This is a big release that supports self-play, rgb, and new versions of SC2, and comes along with a whole lot of other changes. Unfortunately this needed some breaking changes, but they should be fairly easy to fix as well.

Changes:

• Add RGB as an observation and action space
• Add observations for list of units visible in feature layers
• Add observations for last actions, alerts and errors
• Support multi-player/self-play through the environment
• Support different observation types per player when doing self-play
• Support newer versions of the game, up through 4.3
• Support the new actions and units from the 4.0 balance patch
• Better support for launching the right game version for a replay
• Render RGB through the human renderer, including acting with the rendered actions
• Easily generate videos of a replay
• Automatically find the SC2 binary in non-standard setups
• Add enums for units, functions, argument types, and a few observations
• Add an easier way to create FunctionCalls
• Be able to reference bits of the observation by name
• Script to run all the tests
• Add a mock environment for easier testing
• Start testing types with pytype
• Better error messages
• Support more actions in the renderer
• Better rendering of overlays (actions, game speed, selection, build queue)
• Better IPv6 support
• Framework for tracking metrics of steps and episodes
• Allow disabling fog of war
• Support maps from newer seasons.

Breaking changes:

• Feature layer observations are renamed from "screen" and "minimap" to "feature_screen" and "feature_minimap" to differentiate from "rgb_screen" and "rgb_minimap"
• Resolution is no longer optional for environment construction
• The way to specify player race and difficulty to the environment has changed
• Races and difficulties are specified as full words, not single letters
• lib/init.py no longer exports any variables. Get them from lib.actions or lib.features directly.
• action and observation specs are now tuples to reflect the fact that actions/observations are multiplayer.

Experimental work:

• Play human vs agent.
• Play remotely, eg: sc2ai.net ladder.

Version 1.2

• Better support for multiple languages.
• Support versions up to 3.19
• Start the right version for a replay
• Use absl for better python3 support
• Fix some bugs in the mini games

Initial release

This is the initial release of PySC2.