ViaMOBO Optimizes High-Dimensional Multi-Objective Black-Box Problems.

Multi-objective Bayesian optimization (MOBO) is a powerful technique for finding optimal trade-offs (Pareto fronts) in expensive, black-box problems, where function evaluations are costly. However, most existing MOBO methods struggle with high-dimensional decision spaces due to their exponentially increasing sampling complexity. This research presents ViaMOBO, a generic framework designed to tackle these high-dimensional challenges. The core innovation of ViaMOBO lies in its use of a variable interaction analysis model. This model determines whether the high-dimensional decision space can be fully or partially decomposed into smaller, independent or interdependent subspaces. By understanding these variable relationships, ViaMOBO can then perform localized Bayesian optimization within these divided subspaces, significantly reducing the overall complexity. This approach allows ViaMOBO to adaptively handle problems where objectives might be separable, partially separable, or non-separable without strong prior assumptions. Comparative experiments against state-of-the-art MOBO methods on both synthetic and real-world benchmarks demonstrated that ViaMOBO consistently outperformed its baselines in accurately approximating the Pareto front for high-dimensional, expensive multi-objective problems.