This is the repo for the PKU 2023 CoRe Submission Decision Transformer for Modeling Theory of Mind in Multi-Agent Systems.
Our code is based on the Overcooked-AI environment.
It is useful to setup a conda environment with Python 3.7 (virtualenv works too):
conda create -n overcooked_ai python=3.7
conda activate overcooked_ai
Clone the repo
git clone https://github.com/namespacebilibili/overcooked_ai.git
Install requirements:
conda install --file requirements.txt
And
pip install -e overcooked_ai
To generate dataset for DT, run
cd overcooked_ai/src/human_aware_rl/imitation && python dt_dataset.py
Train DT: run
cd overcooked_ai/src/human_aware_rl/imitation && python train_dt.py
To reproduce the BC result, run
cd overcooked_ai/src/human_aware_rl/imitation && python reproduce_bc.py
To train self-play PPO, run
cd overcooked_ai/src/human_aware_rl/ppo && ./run_experiment.sh
To evaluate our agent, run the corresponding function at overcooked_ai/src/human_aware_rl/ppo/evaluate.py.