Preference-Shaped Expected Hypervolume and R2 Improvement: Exact Computation and Monotonicity

This academic paper advances Bayesian multiobjective optimization by clarifying how preference transformations affect two key performance indicators—hypervolume and R2—used in algorithm design. The research provides exact computational methods and proves that R2 improvement, contrary to prior assumptions, cannot be directly computed as objective-space hypervolume but instead represents volume in scalarization space, enabling new algorithmic implementations.

This paper addresses a fundamental theoretical gap in multiobjective optimization, a field increasingly relevant to machine learning and decision-making systems. The authors systematically distinguish between two geometric frameworks—hypervolume indicators based on dystopian reference points versus R2 indicators using utopian points—that practitioners often treat interchangeably despite their structural differences. The research carries implications for algorithm designers implementing Bayesian optimization across multiple competing objectives, a scenario common in hyperparameter tuning, resource allocation, and engineering design problems.

The key insight concerns R2 improvement's true nature: while researchers might expect it to align with weighted hypervolume computations in objective space, the paper proves an obstruction exists at lower dimensions. The Lebesgue-density hypervolume representation cannot capture certain boundary contributions that Tchebycheff scalarization methods detect. This discovery fundamentally reshapes how practitioners should approach exact computation of R2 improvement, shifting focus from objective space to scalarization space where the volume measurement becomes tractable.

The practical impact extends to algorithm implementation. By representing exact integral R2 improvement as Tchebycheff shadow volume, the authors enable three concrete computational approaches: finite-sum algorithms for discrete settings, quadrature methods for continuous optimization, and Gaussian surrogate formulations where R2 improvement decomposes into scalar expected improvements. This framework eliminates previous computational ambiguities and provides engineers with principled methods for building efficient multiobjective Bayesian optimization systems, particularly valuable in machine learning contexts where multiple objectives frequently conflict.

• →Hypervolume and R2 indicators, despite similar algorithmic roles, possess fundamentally different geometric properties requiring distinct computational approaches.
• →R2 improvement cannot be directly computed as objective-space hypervolume due to lower-dimensional boundary effects that Lebesgue density cannot capture.
• →Exact R2 improvement equals the scalarization-space volume between Tchebycheff envelopes, enabling practical finite-sum and quadrature algorithms.
• →The research provides three concrete computational methods: discrete finite-sum, continuous quadrature, and Gaussian surrogate formulations.
• →Preference transformations must be carefully analyzed to preserve exactness, Pareto compatibility, and monotonicity in multiobjective optimization.