Efficient Multi-agent Communication via Shapley Message Value

This is the official implementation for our paper Efficient Multi-agent Communication via Shapley Message Value, based on the open-sourced codebases PyMARL2 which is built on the codebase PyMARL.

Installation instructions

conda create -n sms python==3.7
conda activate sms
pip install -r requirements.txt

Set up listener-speaker environment, which is developed based on the open-sourced repo lb-foraging:

cd listener_speaker
pip install -e .

Run an experiment

python3 src/main.py --config=[Algorithm name] --env-config=[Env name] with [key1=value1] [key2=value2] ...

To run an exiperiment, you need to provide the config files for both the algorithm and environment. These files act as default configurations for them and are located in src/config. --config refers to the config files in src/config/algs, which includes SMS and some other baselines. --env-config refers to the config files in src/config/envs. You can mofidy the hyper-parameters for an experiment by specifying them after the with keyword. Once an experiment is done, the results will be stored in the results folder.

For example, to run SMS on the listener-speaker environment with $4$ listener-speaker pairs on a $15\times 15$ grid world:

python3 src/main.py \
    --config=sms \
    --env-config=listener_speaker \
    with \
    env_args.key="listener_speaker:ListenerSpeaker-15x15-4-v1" \
    msg_dim=2 \
    entropy_coef=0.0 \
    start_msg_select_timestep=200000 \

Run SMS on the task hallway:

python3 src/main.py \
    --config=sms \
    --env-config=hallway \
    with \
    start_msg_select_timestep=300000 \

We also provide corresponding scripts to run the experiments:

sh scripts/run_ls_sms.sh
sh scripts/run_hw_sms.sh