ViaMOBO: A New Framework for High-Dimensional Multi-Objective Bayesian Optimization
A recent paper published on arXiv (2608.11713) presents ViaMOBO, a versatile framework aimed at tackling costly multi-objective black-box optimization challenges characterized by high-dimensional decision spaces. This innovative method employs an analysis of decision variable interactions to assess whether the decision space can be entirely or partially segmented, thus facilitating local Bayesian optimization within the derived subspaces. This technique effectively mitigates the exponential sampling complexity that restricts existing MOBO approaches to low-dimensional scenarios. The variable interaction model discerns the separability of objectives—whether they are separable, partially separable, or non-separable—based on the interrelations among decision variables, without imposing stringent assumptions. While the authors are not specified in the abstract, this work is pivotal for enhancing MOBO in high dimensions, with implications for engineering design, hyperparameter tuning in machine learning, and other domains involving the optimization of expensive black-box functions.
Key facts
- Paper arXiv:2608.11713 introduces ViaMOBO, a framework for high-dimensional multi-objective Bayesian optimization.
- ViaMOBO uses decision variable interaction analysis to divide the decision space.
- It performs local Bayesian optimization in divided subspaces.
- The method can determine if objectives are separable, partially separable, or non-separable.
- It addresses exponential sampling complexity in high-dimensional MOBO.
- The paper is a cross-type announcement on arXiv.
- The approach is generic and does not rely on strong assumptions.
- Potential applications include expensive black-box problems in various fields.
Entities
Institutions
- arXiv