Meta-Black-Box optimization for evolutionary algorithms: Review and perspective

Abstract

Black-Box Optimization (BBO) is increasingly vital for addressing complex real-world optimization challenges, where traditional methods fall short due to their reliance on expert knowledge and time-consuming processes. Meta-Black-Box Optimization (MetaBBO) emerges as a pivotal solution, leveraging meta-learning to enhance or discover optimization algorithms automatically. Originating from Automatic Algorithm Design (AAD), MetaBBO has branched into areas such as Learn to Optimize (L2O), Automated Design of Meta-heuristic Algorithm (ADMA), and Automatic Evolutionary Computation (AEC), each contributing to the advancement of the field. This comprehensive survey integrates and synthesizes the extant research within MetaBBO for Evolutionary Algorithms (EAs) to develop a consistent community of this research topic. Specifically, a mathematical model for MetaBBO is established, and its boundaries and scope are clarified. The potential optimization objects in MetaBBO for EAs is explored, providing insights into design space. A taxonomy of MetaBBO methodologies is introduced, reflecting the state-of-the-art from a meta-level perspective. Additionally, a comprehensive overview of benchmarks, evaluation metrics, and platforms is presented, streamlining the research process for those engaged in learning and experimentation in MetaBBO for EA. The survey concludes with an outlook on research, underscoring future directions and the pivotal role of MetaBBO in automatic algorithm design and optimization problem-solving.