Changed

• Split out subjects and add more checks for VideoSubject
• Update supported version of opencv

Fixed

• #246
  • Fix clean_up_20 to use the correct puppet bot
  • Delete territory__rooms_6 as it is a duplicate of territory__rooms_5

Update Melting Pot version to 2.3.1.

Added

• The results of the Melting Pot Competition, NeurIPS 2023.
• Updates to the evaluation notebook to use the new results.

[2.3.0] - 2024-06-27

Added

• New scenarios, puppets and bots used in the Melting Pot Contest, NeurIPS 2023.

Known issues

• Two scenarios in the competition were set up wrongly. Tracking issue:
  #246
  • clean_up_20 is using the wrong (puppet) bot:
    clean_up__puppet_sanctioning_alternator_nice_0 instead of
    clean_up__puppet_sanctioning_alternator_0.
  • territory__rooms_6 is identical to territory__rooms_5.

Fixed

• Update setup.py to work with earlier setuptools (fixes broken 2.2.1 release).

Changed

• Do not allow orientation = None in Transform objects
• Improve documentation for territory__*.
• Improve scenario tags.
• Move puppeteers testutils to testing dir.
• remove restriction on chex
• Update setup.py to work with older versions of setuptools
• Add dev tools to setup.py
• Add pytest-xdist as a required plugin.
• Update pylintrc

Fixed

• Debug observations break hidden agenda #168
• Various lint errors

Removed

• Residual v1 file: reaction_graph_utils.py
• Remove stale documentation left over from 1.0.

Changed

• Make meltingpot pip installable.
• Support import meltingpot as an alias of import meltingpot.python.
• Raise minimum Python version to Python 3.10.
• Install assets as part of pip install .
• Update Dockerfile and dev tooling (pytest, pyink, pytype, pylint, isort).
• Update examples to work with v2.
• Update README.md with new installation details.
• Migrate from rx to reactivex.

Fixed

• Remove type annotation for dtype.
• Use correct roles in play_hidden_agenda.
• Fix noop that was causing a typing error.
• Add missing __init__.py files.
• Set the default orientation to NORTH for objects that have no orientation
  defined.

Added

• Evaluation utilities.

Changed

• Added COLLECTIVE_RETURN to PERMITTED_OBSERVATIONS.
• Split install.sh into three scripts.
• Move from yapf to black.
• Remove debug observations to speed up environment stepping.

Fixed

• Add Lua 5.2 compatibility for unpack.

Removed

• SubstrateWrappers previously needed for the v1 bots.

Added

• Colab for visualizing evaluation results.
• Tests for the examples.
• Mocks of specific substrates and scenarios, for use in testing.
• Helper for setting world.rgb spec.

Changed

• Improve debugging information in SavedModelPolicy.
• Resample bots at the beginning of scenario episodes.

Fixed

• Initialize Transform before any other component (#84, #24).

Added

• New substrate "Hidden Agenda" and its scenarios.

Melting Pot Version 2.0.0 release. See Melting Pot 2.0 Tech Report for detailed information on the new substrates, bots, and scenarios.

Changed

• Removed all v1 scenarios, bots, and substrates and replaced with new versions.
• Scenarios now support heterogenous roles, which must be specified at build
  time.
• Various improvements to examples and their documentation.

Added

• New puppeteers and policies to implement new bots.
• New utils to handle sprites, colors, and maps.
• Mocks for use in testing.

Changed

• Drop support for Python 3.7 and 3.8.
• Store saved models on Google Cloud Platform rather than in the Git repo.
• x2 speed improvement to some substrates
• Improved install.sh script and installation documentation.
• Various improvements to examples and their documentation.

Fixed

• Puppets were sharing state in Scenarios. #70
• Various issues with RLlib examples.

Added

• .devcontainer for the project.
• pettingzoo example.
• TF1-compatible version of SavedModelPolicy.