Accurate Peak Detection in Multimodal Optimization via Approximated Landscape Learning

Here we provide sourcecodes of APDMMO, which is accepted by GECCO 2025.

Citation

@article{ma2025accurate,
  title={Accurate Peak Detection in Multimodal Optimization via Approximated Landscape Learning},
  author={Ma, Zeyuan and Lian, Hongqiao and Qiu, Wenjie and Gong, Yue-Jiao},
  journal={arXiv preprint arXiv:2503.18066},
  year={2025}
}

Run the code

The optimization process on CEC2013 multimodal optimization benchmark can be activated via the command below.

python run.py

Results

The optimization process includes three stages:

1. Global Landscape Fitting(GLF): This stage is used to train a Landscape Learner for a problem. The Log files will be saved to ./log, and the trained models will be saved to ./checkpoint, while the information of mu and std for normalization are stored in ./mu_std_info. The contour maps of 1D/2D problems are saved in ./pic.

2. Free-of-trial Peak Detection(FPD): This stage is used to detect potential peak areas. The optimized solutions are saved in ./optimization_result.

3. Parallel Local Search(PLS): The last stage is designed to perform local optimization. The results of Peak Ratio(PR) and Success Rate(SR) are saved to ./result.