This codebase accompanies the paper submission "LAGMA: LAtent Goal-guided Multi-Agent Reinforcement Learning" and is based on PyMARL and SMAC which are open-sourced. The paper is accepted by ICML2024.
PyMARL is WhiRL's framework for deep multi-agent reinforcement learning.
To train LAGMA on SC2 (dense reward) setting tasks, run the following command:
python3 src/main.py --config=lagma_sc2 --env-config=sc2 with env_args.map_name=5m_vs_6m
To train LAGMA on SC2 (sparse reward) setting tasks, run the following command:
python3 src/main.py --config=lagma_sc2_sparse_3m --env-config=sc2_sparse with env_args.map_name=3m
To train LAGMA on GRF (sparse reward) setting tasks, run the following command:
python3 src/main.py --config=academy_3_vs_1_with_keeper --env-config=lagma_grf_3_vs_1WK
If you find this repository useful, please cite our paper:
@inproceedings{na2024lagma,
title={LAGMA: LAtent Goal-guided Multi-Agent Reinforcement Learning},
author={Na, Hyungho and and Moon, Il-chul},
booktitle={The Forty-first International Conference on Machine Learning},
year={2024}
}