Multi-agent Combat Arena (MaCA) is a heterogeneous multi-agent distributed decision and control technology reasearch platform produced by CETC-TFAI team. It focuses on the application of AI technologies e.g. reinforcement learning in multi-agent cooperation and confrontation

• Linux 64-bit or Mac OS with Python 3.6
• numpy 1.14.2 or later
• pygame 1.9.3 or later

There is no limitation on agents' structure. You can write rule-based algorithms or use deep learning frameworks.

pip install numpy pygame
git clone https://github.com/CETC-TFAI/MaCA.git
cd MaCA
export PYTHONPATH=$(pwd)/environment:$PYTHONPATH

fight.py can execute two agents. It uses two instances of a fixed-rule agent to fight each other by default.

python fight.py

You can specify agents and map by input arguments. In addition an agent should provide a call interface follows the MaCa platform specification.

MaCa can record runtime log while playing and training. Use replay.py to perform a replay.

First, run fight.py and enable log record function

python fight.py --log

Then, run replay.py to replay the log

python replay.py default_log

The log structure of MaCa is a set of .macalog files, they will save in path "log/log-name/". When you run the replay.py, You must input a "log-name" as the parameter to specify which log you want to replay.

python train.py

MaCA provides an image based RL-API for deep reinforcement learning, but also a underlying data based RAW-API for other agent types e.g. fixed-rule agent development. Considering that different algorithms need different implementation methods and code structures, the training example can not be suitable for all algorithms. You can design your own algorithm and training script based on the two APIs.