Prism: Spectral Parameter Sharing for Multi-Agent Reinforcement Learning

This is the official implementation of our paper
"Prism: Spectral Parameter Sharing for Multi-Agent Reinforcement Learning."

---

Getting Started

1. Pull Docker Image

We provide a pre-configured Docker image on Docker Hub:

docker run -it --gpus all \
  --name Prism \
  --user root \
  --net host \
  -v /path/to/your/code:/app \
  -v /path/to/your/StarCraftII:/data \
  --ipc host \
  kkgb/prism:v1 bash

2. Conda

# for Level-based Foraging
conda activate LBF

# for SMACv2
conda activate SMACv2

# for MaMuJoCo
conda activate MaMuJoCo

3. Run

For single run

# for Level-based Foraging
python src/main.py --config=Prism --env-config=gymma --exp-config=lbf-10x10-3p-3f --seed 0

# for SMACv2
python src/main.py --config=Prism --env-config=sc2_gen_terran --exp-config=terran_5v5_10M_s0 seed=0

# for MaMuJoCo
python examples/train.py --load_config=tuned_configs/Prism/Ant-v2-4x2-Prism.json --seed 0

For Multiple run

# for Level-based Foraging
bash run_docker.sh \
  Prism \
  <maps> \
  <num_runs> \
  <gpu_ids> \
  <num_repeats> \
  <seeds> \
  /path/to/your/code \
  /path/to/your/StarCraftII \
  <wandb_api_key>

# for SMACv2
bash run_docker.sh \
  Prism \
  <env_config> \
  <exp_config> \
  <num_runs> \
  <gpu_ids> \
  <num_repeats> \
  <seeds> \
  /path/to/your/code \
  /path/to/your/StarCraftII \
  <wandb_api_key>

# for MaMuJoCo
bash run_docker.sh \
  Prism \
  <maps> \
  <num_runs> \
  <gpu_ids> \
  <seeds> \
  /path/to/your/code \
  /path/to/your/StarCraftII \
  <wandb_api_key>

For detailed setup, please refer to the README.md file in each corresponding folder.

Citing

@article{kim2026prism,
  title={Prism: Spectral Parameter Sharing for Multi-Agent Reinforcement Learning},
  author={Kim, Kyungbeom and Oh, Seungwon and Kim, Kyung-Joong},
  journal={arXiv preprint arXiv:2602.06476},
  year={2026},
  url={https://arxiv.org/abs/2602.06476}
}