Left: Short-horizon MOSMAC scenarios (example: 4t); Right: Long-horizon MOSMAC scenarios with complex terrains

NEWS: MOSMAC has been accepted at AAMAS 2025 as a full paper. Please cite the full paper with the following BibTex:

@inproceedings{geng_mosmac_2025,
    abbr={AAMAS 2025},
    address = {Richland, SC},
    series = {AAMAS'25},
    title = {MOSMAC: A Multi-agent Reinforcement Learning Benchmark on Sequential Multi-objective Tasks},
    booktitle = {Proceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems},
    publisher = {International Foundation for Autonomous Agents and Multiagent Systems},
    author = {Geng, Minghong and Pateria, Shubham and Subagdja, Budhitama and Tan, Ah-Hwee},
    month = may,
    year = {2025},
    html={https://aamas2025.org/index.php/conference/program/accepted-papers/},
    }

MOSMAC provides a set of multi-objective multi-agent reinforcement learning (MOMARL) tasks extending from SMAC, which originally focused on single-objective multi-agent reinforcement learning (MARL) tasks in the StarCraft II environment.

Specifically, MOSMAC includes the following features:

• Multi-Objective Tasks: Reinforcement learning agents in MOSMAC scenarios need to learn policies that simultaneously balance multiple objectives beyond combat. Specifically, MOSMAC considers objectives including:
  • Combat: a widely adopted objective in SC2 environment, originally presented by Samvelyan et al. in SMAC (2019),
  • Safety: also know as escape by Hu et al. in MO-MIX (2023),
  • Navigate: a novel objective where agents need to navigate to a target location, firstly presented in our AAMAS paper (2024).
• Sequential Task Allocation: As a closer simulation to real-world scenarios, MOSMAC includes a set of scenarios that challenge agents with sequential task allocation, where agents need to complete multiple tasks sequentially.

Users should execute commands from the root directory with Python 3 to start training/evaluation processes.

Following is an example of training the IQL algorithm on the 3t scenario with a total running step of 2,050,000:

python3 src/main.py --config=iql --env-config=shcfc_beta with env_args.map_name=3t t_max=2050000

Following is an example of training MADDPG on the 4t_vs_4t task with complex terrain features and a total running step of 10,050,000:

python3 src/main.py --config=maddpg --env-config=lhcfcws with env_args.map_name=4t_vs_4t_large_complex env_args.final_target_index=13 env_args.obs_pathing_grid='True' cuda_id='cuda:0' t_max=10050000

In the above examples, the config option specifies the configuration of the selected MARL algorithm, the env-config option specifies the environment, and the map_name states the task(map) for training. For the MOSMAC scenarios with single-task settings, users should select shcfc_beta as the env-config option. For the MOSMAC scenarios with sequential task allocation, users should select lhcfcws as the env-config option.

If you use MOSMAC in your research, please cite our AAMAS'24 paper: Benchmarking MARL on Long Horizon Sequential Multi-Objective Tasks.

Minghong Geng, Shubham Pateria, Budhitama Subagdja, and Ah-Hwee Tan. 2024. Benchmarking MARL on Long Horizon Sequential Multi-Objective Tasks. In Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems (AAMAS '24). International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 2279–2281.

In BibTeX format:

@inproceedings{10.5555/3635637.3663133,
    author = {Geng, Minghong and Pateria, Shubham and Subagdja, Budhitama and Tan, Ah-Hwee},
    title = {Benchmarking MARL on Long Horizon Sequential Multi-Objective Tasks},
    year = {2024},
    isbn = {9798400704864},
    publisher = {International Foundation for Autonomous Agents and Multiagent Systems},
    address = {Richland, SC},
    booktitle = {Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems},
    pages = {2279–2281},
    numpages = {3},
    location = {<conf-loc>, <city>Auckland</city>, <country>New Zealand</country>, </conf-loc>},
    series = {AAMAS '24}
}

MOSMAC is implemented the Extended PyMARL (EPyMARL) framework. If you use EPyMARL in your research, please cite the following paper: Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks

Georgios Papoudakis, Filippos Christianos, Lukas Schäfer, & Stefano V. Albrecht. Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks, Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS), 2021

In BibTeX format:

@inproceedings{papoudakis2021benchmarking,
   title={Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks},
   author={Georgios Papoudakis and Filippos Christianos and Lukas Schäfer and Stefano V. Albrecht},
   booktitle = {Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS)},
   year={2021},
   url = {http://arxiv.org/abs/2006.07869},
   openreview = {https://openreview.net/forum?id=cIrPX-Sn5n},
   code = {https://github.com/uoe-agents/epymarl},
}