Multi-Objective Kernel Ridge Regression-Assisted Subproblem Co-Optimization for Expensive Multi-Objective Binary Optimization
针对昂贵多目标二元优化问题，提出一种基于分解框架的方法，集成多目标核岭回归统一代理模型、子问题协同优化和多样性驱动采样准则，在有限评估预算下提升收敛性和多样性。
Expensive multi-objective binary optimization problems frequently emerge in real-world applications, where evaluating a single solution incurs significant computational or physical costs. The combinatorial nature of the binary search space further exacerbates the difficulty, often resulting in poor scalability and slow convergence for traditional multi-objective evolutionary algorithms. Although surrogate-assisted multi-objective evolutionary algorithms provide a promising alternative by leveraging models to predict the utility of candidate solutions or provide search guidance, most existing approaches model each objective independently and struggle to maintain a proper balance between convergence and diversity under strict evaluation budgets. To address these challenges, we propose a novel method within the decomposition-based framework: Multi-objective Kernel Ridge Regression-Assisted Subproblem Co-optimization (MOKRR-SCo). The algorithm integrates three key components: 1) Multi-objective Kernel Ridge Regression (MOKRR), which constructs a unified surrogate model using a shared weighted Hamming kernel to jointly approximate multiple objectives; 2) MOKRR-driven Subproblem Co-optimization (MSCO), which leverages historical data and MOKRR model predictions to collaboratively guide search across subproblems; and 3) Diversity-Driven Infill Sampling Criterion (DDISC), which selects sparse solutions from the MSCO-evolved population for expensive evaluations for each subproblem, enhancing diversity and mitigating the adverse effects of surrogate errors. Extensive experiments on benchmark and real-world expensive multi-objective binary optimization problems demonstrate that MOKRR-SCo consistently outperforms state-of-the-art algorithms in both convergence and diversity under tight evaluation budgets. The source codes of the proposed MOKRR-SCo algorithm are available from.