A Multi-Agent Reinforcement Learning survival environment in 2D with a pseudo--gym interface.
See here for details on the environment semantics.
This is not a proper Python package yet; python source code is in the
masurvival directory.
The code assumes that the version 3.9 of Python is used. The dependencies are those listed in requirements.txt and PyBox2d 2.3.10, which can be installed with conda from the conda-forge channel. More details on installing dependencies can be found below.
The environment itself (without rendering) only depends on Gym, Numpy and PyBox2d. Rendering depends only on PyGame. The demo.py script depends on pygifsicle and imageio for screenshot and gif recording.
The most convenient way to install all dependencies is using conda, which is needed to install PyBox2d anyway. It is convenient to create a dedicated conda environment with
conda create -n env-name python=3.9
conda activate env-name
Since PyBox2d must be installed from conda-forge and gym fails to install with the conda-forge installation of pip, the preferred way to install all the dependencies using conda is to append the conda-forge channel to the conda config, so that the defaults channel is used for packages other than PyBox2d. This can be done with
conda config --append channels conda-forge
The packages in requirements.txt can then be installed using pip by running
pip install -r requirements.txt
PyBox2d can then be installed with conda with
conda install pybox2d