Sequential Multi-agent Dynamic Algorithm Configuration

Official implementation of NeurIPS 2025 paper Sequential Multi-Agent Dynamic Algorithm Configuration

Installation

conda create -n madac python=3.7.13
conda activate madac
pip install -r requirements.txt
pip install -e . # Local installation of madacbench Package

Train Seq-MADAC

Train Seq-MADAC with MOEA/D in different tasks.

# Train
python algos/madac/main.py --config=sadn_ns --env-config=moea with env_args.key=M_2_46_3

Train Seq-MADAC with 10 D Seq-Sigmoid-Mask.

python algos/madac/main.py --config=sadn_sigmoid --env-config=sigmoid_state with env_args.mask=True env_args.n_agents=10

You can modify the relevant configuration file algos/madac/config/envs/moea.yaml and algos/madac/config/algs/sadn_ns.yaml. And you can train Seq-MADAC with different algorithms by specifying algorithm in the config (e.g., using VDN by --config=vdn_ns)

Test Seq-MADAC

The trained model is saved in directory results/madac/models/, you need to specify the model directory in the configuration file via parameter checkpoint_path. More details of the configuration file can be found in EPyMARL.

Test Seq-MA-DAC in a specific problem. (The problem set is DTLZ2_3 DTLZ4_3 WFG4_3 WFG5_3 WFG6_3 WFG7_3 WFG8_3 WFG9_3 ) For example,

python algos/madac/main.py --config=sadn_ns_test --env-config=moea_test with env_args.key=DTLZ2_3

Other Baselines

MOEA/D

python algos/moead/moead_baseline.py

DQN

Train DQN in different tasks. (The task set is M_2_46_3, M_2_46_5, M_2_46_7). For example,

python algos/dac/dqn.py --key M_2_46_3

The trained model can be found in the directory results/dqn/M_2_46_3

The command to test the corresponding model on all problems is

python algos/dac/test_dqn.py --key M_2_46_3

License

All the source code that has been taken from the EPyMARL repository was licensed (and remains so) under the Apache License v2.0 (included in LICENSE file). Any new code is also licensed under the Apache License v2.0.

Citation

@inproceedings{seq-madac,
    author = {Chen Lu, Ke Xue, Lei Yuan, Yao Wang, Yaoyuan Wang, Fu Sheng, Chao Qian},
    title = {Sequential Multi-Agent Dynamic Algorithm Configuration},
    booktitle = {Advances in Neural Information Processing Systems 38 (NeurIPS'25)},
    year = {2025},
    address={San Diego, CA}
}