昂贵的多目标优化问题广泛存在于现实世界中的不同应用领域。其优化目标通常为黑盒函数，且求得其真实目标函数值的评估代价高昂；而受限于现实世界的有限资源和成本，求解器只允许进行有限次函数评估用于搜索该类问题的帕累托前沿。多目标贝叶斯优化方法能够有效地求解该类问题，其利用高斯过程代理模型近似原优化问题降低函数评估成本，并利用平衡利用和探索之间关系的获取函数推荐候选解。本文针对低维和高维决策空间中的并行化函数评估问题、获取函数优化效率问题以及维度灾难和边界问题，对多目标贝叶斯优化方法进行了下述四个方面的研究，旨在有效地求解低维和高维昂贵的多目标优化问题。在低维决策空间中，针对串行化函数评估不能充分利用并行硬件计算资源的问题，本文提出了\textbf{基于自适应采样的批量多目标贝叶斯优化方法}。其通过双目标获取函数和自适应选点策略，实现了并行化函数评估和搜索过程中收敛性与多样性的动态平衡，使得最终近似帕累托前沿更加均匀地分布在整个搜索空间。在高维决策空间中，针对当前多目标贝叶斯优化方法面临的缺乏通用性框架以及维度灾难和边界问题，本文首先提出了\textbf{基于块坐标更新的高维多目标贝叶斯优化方法}，利用块坐标更新策略和嵌入帕累托支配关系的获取函数，降低决策空间维度并提高获取函数的优化效率，从而缓解了维度灾难和边界问题。为了避免上述方法中决策空间划分的随机性和串行化函数评估，本文进一步提出了\textbf{基于可加高斯结构的高维多目标贝叶斯优化方法}。其通过引入先验知识避免了决策空间划分的随机性；利用可加高斯结构和可加双目标获取函数实现了决策空间降维和高维决策空间中的并行化函数评估，提高了获取函数的优化效率。然而，先验知识使得决策空间划分具有一定主观性，且该方法假设优化目标可分，导致其假设性过强、泛化能力低。为了解决上述问题，本文提出了\textbf{基于变量交互分析的高维多目标贝叶斯优化方法}。其通过分类器学习交互变量分组用于判断优化目标是否可分；并利用可加多目标获取函数和虚拟导数信息提高了获取函数优化效率。标准多目标测试问题和交通领域优化问题上的实验结果表明，本文关于低维决策空间的研究实现了并行函数评估、很好地平衡了收敛性与多样性；关于高维决策空间的研究缓解了维度灾难和边界问题的同时，提高了获取函数的优化效率。

Expensive multi-objective optimization problems (MOPs) arise in a wide variety of real-world applications. They are usually black-box functions, and evaluating the true function values is costly. Due to limited hardware resources and high evaluation costs in the real world, only a limited number of function evaluations can be conducted when searching for the Pareto fronts of such problems. Multi-objective Bayesian optimization (MOBO) methods can effectively solve this type of problem by using a Gaussian process surrogate to approximate the original MOP, reducing the function evaluation cost, and employing an acquisition function that balances exploitation and exploration to recommend candidate solutions. This article focuses on parallel function evaluation, the optimization efficiency of the acquisition function, the curse of dimensionality and boundary issues in low-dimensional and high-dimensional decision spaces, and carries out the following four studies, aiming at effectively solving the expensive MOPs.In low-dimensional decision space, this article proposes \textbf{An Adaptive Batch Multi-objective Bayesian Optimization method, Adaptive Batch-ParEGO} to address the challenge of sequential function evaluation that cannot fully utilize parallel hardware computing resources. By using a bi-objective acquisition function and an adaptive selection strategy, this method achieves parallel function evaluation and dynamically balances exploitation and exploration during the search process. As a result, the final approximate Pareto front is more evenly distributed throughout the entire search space.In high-dimensional decision space, existing multi-objective Bayesian optimization methods face several challenges, including the lack of a general framework, the curse of dimensionality, and boundary issues. To address these challenges, this article proposes \textbf{High-dimensional Multi-objective Bayesian Optimization Method with Block Coordinate Updates, Block-MOBO}. It uses block coordinate update and an acquisition function embedded with the Pareto dominance relationship to reduce the decision space dimensionality and improve the optimization efficiency of the acquisition function, thus alleviating the curse of dimensionality and boundary issues. To avoid the randomness of decision space partition and sequential function evaluation in the above method, this article further proposes \textbf{High-dimensional Multi-objective Bayesian Optimization Method with Additive Gaussian Structure, ADD-HDMBO}. This method introduces prior knowledge to avoid the randomness in decision space partition and uses additive Gaussian structure and a bi-objective acquisition function to achieve decision space dimensionality reduction and parallel function evaluation in high-dimensional decision spaces. This significantly improves the optimization efficiency of the acquisition function. However, prior knowledge makes decision space partition subjective, and the method assumes that optimization objectives are separable, leading to a strong assumption and low generalization ability. To address these issues, this paper proposes \textbf{High-dimensional Multi-objective Bayesian Optimization Method with Variable Interaction Analysis, ViaMOBO}. It uses a classifier to learn interactive variable grouping to determine whether the MOPs are separable. If the MOPs are separable, this method uses an additive multi-objective acquisition function and virtual derivative information to improve the optimization efficiency of the acquisition function.The experimental results on multi-objective test problems and problems in the transportation area demonstrate that this article‘s research on low-dimensional decision space in this article achieves parallel function evaluation and balanced convergence with diversity well. The research on high-dimensional decision space alleviates the curse of dimensionality and boundary issues and significantly improves the optimization efficiency of the acquisition function.