Dual Relational Surrogate-Assisted Evolution for Expensive Constrained Multiobjective Optimization
针对昂贵约束多目标优化问题，提出自适应多阶段进化框架RECMO，利用两个关系代理模型分别引导收敛和可行性，动态切换优化重点，在基准和26个真实问题上优于现有算法。
Expensive constrained multiobjective optimization problems (ECMOPs) present significant challenges due to the high cost of evaluating both objective and constraint functions. Existing constrained multiobjective evolutionary algorithms (CMOEAs) often struggle with convergence and feasibility under strict evaluation budgets, while surrogate-assisted CMOEAs typically rely on regression models that do not align well with the ranking-based nature of evolutionary selection. To address these limitations, this paper proposes RECMO, an adaptive multi-stage evolutionary framework driven by dual relational surrogate models. One model captures dominance relationships to promote convergence, while the other focuses on feasibility comparisons to guide constraint satisfaction. These surrogates are dynamically activated according to the stage of evolution: the convergence-oriented surrogate dominates in early search, the feasibility surrogate is gradually introduced in the middle stage when conflicts arise, and both are jointly leveraged in the late stage to ensure feasible and high-quality solutions. This selective activation enables RECMO to progressively shift the optimization emphasis from objective-driven exploration to feasibility enforcement in a resource-efficient manner. Extensive experiments on benchmark suites and 26 real-world ECMOPs demonstrate that RECMO consistently outperforms state-of-the-art algorithms in terms of convergence, diversity, and feasibility. Additional ablation studies and sensitivity analyses further confirm the robustness and adaptability of the proposed framework.