DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization

This project provides the sourcecodes of DesignX, which has been recently accpeted by NeurIPS 2025

Citation

The PDF version of the paper is available here. If you find our DesignX useful, please cite it in your publications or projects.

@inproceedings{guo2025designx,
  title={DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization},
  author={Guo, Hongshu and Ma, Zeyuan and Ma, Yining and Zhang, Xinglin and Chen, Wei-Neng and Gong, Yue-Jiao},
  booktitle={Proceedings of the 39th Conference on Neural Information Processing Systems},
  year={2025}
}

A sister project we also recommend you to read and refer to is ConfigX, which is the basis of DesignX and the origin of the modular evolutionary operator space.

Requirements

Create the conda environment with python 3.9.18 and torch 2.3.1, then install packages:

conda create -n DesignX python=3.9.18
conda activate DesignX
conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

Train

To train the Agent-1, run:

python main.py

To train the Agent-2, run:

python main.py --train_cfx True

For more adjustable settings, please refer to config.py and config_cfg.py for details.

Recording results: Log files will be saved to ./logs, the file structure is as follow:

logs
|--run_name
   |--logging files
   |--...

The saved checkpoints will be saved to ./outputs, the file structure is as follow:

outputs
|--run_name
   |--epoch-0.pt
   |--epoch-1.pt
   |--...

Rollout

Modify load_path (The checkpoint saving directory, default to be "./outputs"), load_name (The run_name of the target Agent-1 model) and load_epoch (The epoch of the model) in config.py to assign the Agent-1 model.

Modify load_path (The checkpoint saving directory, default to be "./outputs"), load_name (The run_name of the target Agent-2 model) and load_epoch (The epoch of the model) in config_cfg.py to assign the Agent-2 model.

Then run:

python main.py --test

to rollout the assigned Agent-1 & 2 models.