/multi-agent

Primary LanguagePythonMIT LicenseMIT

Multi-agent RL environment, in the style of DeepMind's Gathering game.


Screen capture of the game

python demo.py plays a trained RL agent against a random agent. The RL agent can be re-trained in a couple minutes with python reinforce.py.

The map is customizable. All of the maps from DeepMind's paper are available in the maps directory. To use e.g. the "Basic single-entrance region map" from the paper, pass map_name='region_single_entrance' to the GatheringEnv constructor. To specify your own map, create a text file in the same format as the existing maps, and pass map_name='path/to/my/custom/map' to the GatheringEnv constructor.

Installation

pip install -r requirements.txt