yinjiangjin / TMC

Pytorch implementation of "Succinct and Robust Multi-Agent Communication With Temporal Message Control"

Succinct and Robust Multi-Agent Communication With Temporal Message Control

This repository is the official implementation of "Succinct and Robust Multi-Agent Communication With Temporal Message Control" published in NeurIPS 2020.

📋A video demo is available at: https://tmcpaper.github.io/tmc/

Requirements

📋To run the code, please install the SMAC (StarCraft Multi-Agent Challenge) first, which is available at: https://github.com/oxwhirl/smac, please then follow the instructions to install the StarCraft II client. Download StarCraft II into the 3rdparty folder and copy the maps necessary to run over.

📋This code is developed based on the source code of QMIX, which is available at: https://github.com/oxwhirl/pymarl.

📋 To install the requirements, run:

pip install -r requirements.txt

Training

To train the model(s) in the paper, run this command:

python3 src/main.py --config=xxx --env-config=sc2 with env_args.map_name=xxx

📋--config can be one of the following six options: qmix_3s_vs_4z, qmix_3s_vs_5z, qmix_6h_vs_8z. For example 'qmix_6h_vs_8z' means 6h_vs_8z map with QMIX as the mixing network.

📋--env_args.map_name can be one of the following six options: 3s_vs_4z, 3s_vs_5z, 6h_vs_8z.

Code Explanation

📋All the hyperparameters can be found and modified at: src/config/default.yaml, src/config/algs/*.yaml and src/config/envs/*.yaml. The default parameters are for 3s_vs_4z scenario.

📋The test accuracy at each timestep will be saved in the 'accuracy' folder.

📋The test reward at each timestep will be saved in the 'reward' folder.

📋The agent network can be found at: src/controllers/basic_controller_xxx.py, where xxx is the name of the map. For example, the execution of the agent network for 3s_vs_4z can be found at: src/controllers/basic_controller_3s_vs_4z.py

📋The training of agent networks can be found at: src/learners/q_learner_xxx.py, where xxx is the name of the map. For example, the training of the agent network for 3s_vs_4z can be found at: src/learners/q_learner_3s_vs_4z.py

📋Model can be saved by setting save_model = True in src/config/default.yaml. You can also specify the frequency of the saving by changing the save_model_interval option in the config file.

Execution

📋To execute the model, specifying the path of the saved model by filling the checkpoint_path and setting evaluate: True in the src/config/default.yaml.