Reproduce “DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization” on OpenResearch

About this paper

Hongshu Guo, Zeyuan Ma, Yining Ma, Xinglin Zhang, Wei-Neng Chen, Yue-Jiao Gong

DesignX is an automated deep reinforcement learning framework that jointly generates black-box optimization algorithm workflows and dynamically controls their hyperparameters. It consistently designs optimizers that surpass human-crafted algorithms and other automated methods on diverse synthetic and real-world problems, including protein-docking and UAV path planning.

Key insights

• Generating a correct optimizer workflow appears more critical than dynamically controlling hyperparameters for overall performance, as shown by ablation studies.
• DesignX learns to prioritize specific module types (e.g., DE mutation strategies, population reduction mechanisms) based on problem characteristics like modality and search range.
• The choice of initialization strategy has a limited impact on final performance compared to later, more critical algorithmic modules, a non-trivial finding validated by experiments.

Summary by alphaXiv.