MetaSG-SAEA guides search in expensive multi-objective optimization

Existing Meta-Black-Box Optimization (MetaBBO) methods focus on controlling how optimizers search but largely ignore where to search. A new paper by Yukun Du and colleagues introduces MetaSG-SAEA, a bi-level framework that addresses this gap for expensive constrained multi-objective optimization problems (ECMOPs). At the high level, a meta-policy learns region-level guidance; at the low level, a Surrogate-Assisted Evolutionary Algorithm (SAEA) uses that guidance to focus on promising areas. The authors propose Max-Min Constraint-Calibrated Inequality (MM-CCI), a compact, problem-agnostic region abstraction that maps heterogeneous constraint evaluations to an ordered scalar level, backed by theoretical analysis.

To translate region-level guidance into solution-level priors, the framework uses diffusion-based population initialization. Scalability is achieved via an attention-based state representation that works across varying problem dimensions, population sizes, and numbers of objectives and constraints. Experimental results demonstrate that MetaSG-SAEA consistently outperforms state-of-the-art baselines across diverse benchmarks and exhibits strong generalization to unseen problem distributions. This work represents a significant step toward more intelligent, efficient optimization for real-world engineering and scientific applications where each evaluation is costly.