TRAMA

(Official) PyTorch implementation for Trajectory-Class-Aware Multi-Agent Reinforcement Learning (ICLR 2025)

Note

This codebase accompanies the paper submission "Trajectory-Class-Aware Multi-agent Reinforcement Learning (TRAMA)" and is based on PyMARL, SMAC, and SMAC2 which are open-sourced. The paper is accepted by ICLR2025 and now available in OpenReview.

PyMARL is WhiRL's framework for deep multi-agent reinforcement learning and our code includes implementations of the following algorithm:

• LAGMA: LAtent Goal-guided Multi-Agent Reinforcement Learning

Run an experiment

To train TRAMA on surComb3 in SC2(v2), run the following command:

python3 src/main.py --config=trama_gc_qplex --env-config=sc2_gen_protoss_surComb3

To train TRAMA on sc2_gen_protoss in SC2(v2), run the following command:

python3 src/main.py --config=trama_gc_qplex --env-config=sc2_gen_protoss

Publication

If you find this repository useful, please cite our paper:

@inproceedings{na2025trama,
  title={Trajectory-class-aware Multi-agent Reinforcement Learning},
  author={Na, Hyungho and Lee, Kwanghyeon and Lee, Sumin and Moon, Il-chul},
  journal={The Thirteenth International Conference on Learning Representations},
  year={2025}
}