自动黑箱优化算法设计：进展与挑战

초록

As optimization problems in engineering and scientific computing continue to increase in scale and complexity, the traditional paradigm of manually crafting optimizers based on expert knowledge is increasingly constrained by high development costs and limited generalization. Meta-Black-Box Optimization (MetaBBO) is an emerging meta-learning-driven paradigm that seeks to automate black-box optimizer design by learning transferable strategies from a distribution of problem instances, and it has attracted growing attention in intelligent optimization and evolutionary computation. This survey first formalizes the core concepts of MetaBBO and its bi-level optimization framework. It then provides a systematic review of five key meta-tasks from two complementary viewpoints: the optimizer space, including algorithm selection, configuration, composition, and generation; and the decision space, focusing on solution manipulation. We further discuss opportunities and challenges in integrating large language models into MetaBBO, emphasizing their promise and current limitations for enabling end-to-end optimization workflows. Finally, in light of existing gaps in problem understanding, model construction, and training mechanisms, we outline future directions such as building more diverse problem suites, developing more efficient learning paradigms, and advancing hybrid foundation models for multitask meta-optimization. These efforts aim to improve the generality, automation, flexibility, and practical impact of MetaBBO, thereby providing stronger algorithmic support for solving complex real-world optimization problems. © 2026 Science Press. All rights reserved.

키워드