📄 中文摘要
连续黑盒优化领域基准测试面临现有测试套件（如BBOB）结构多样性有限的挑战。探索了将大型语言模型（LLM）嵌入进化循环中，以设计具有明确定义高层景观特性的优化问题的方法。利用LLaMEA框架，引导LLM根据目标属性的自然语言描述生成问题代码。这些目标属性包括但不限于多模态性（例如，指定局部最优点的数量和深度）、可分离性、对称性、平滑度以及整体复杂性。LLM被提示生成Python代码片段，这些片段能够构建具有这些指定特性的目标函数。通过迭代地评估LLM生成的问题代码的性能和特性，并将其反馈给LLM进行优化和调整，形成一个闭环的进化设计过程。这种方法旨在克服传统问题生成方法在灵活性和多样性方面的不足，允许研究人员根据特定的测试需求，快速、有效地生成具有定制化复杂度和结构特性的优化问题。实验结果表明，该框架能够成功生成满足预设高层特性的优化问题，显著提升了基准测试的多样性和针对性，为连续优化算法的开发和评估提供了更丰富的测试环境。
📄 English Summary
LLM Driven Design of Continuous Optimization Problems with Controllable High-level Properties
Benchmarking in continuous black-box optimization is hampered by the limited structural diversity of existing test suites like BBOB. Investigated is the utility of large language models (LLMs) embedded within an evolutionary loop to design optimization problems endowed with clearly defined high-level landscape characteristics. Leveraging the LLaMEA framework, an LLM is guided to generate problem code from natural-language descriptions of target properties. These target properties encompass, but are not limited to, multimodality (e.g., specifying the number and depth of local optima), separability, symmetry, smoothness, and overall complexity. The LLM is prompted to produce Python code snippets capable of constructing objective functions exhibiting these specified attributes. This process involves iteratively evaluating the performance and characteristics of the LLM-generated problem code, feeding this feedback back to the LLM for refinement and optimization, thereby forming a closed-loop evolutionary design process. This methodology aims to overcome the limitations of traditional problem generation methods in terms of flexibility and diversity, enabling researchers to rapidly and effectively generate optimization problems with customized complexity and structural properties tailored to specific testing requirements. Experimental findings demonstrate the framework's success in generating optimization problems that satisfy predefined high-level characteristics, significantly enhancing the diversity and specificity of benchmarking, and providing a richer testing environment for the development and evaluation of continuous optimization algorithms.