General-kernel optimality for cost-sensitive Gaussian-process algorithms

Establish general-kernel optimality guarantees for costly-observation Gaussian-process algorithms under observation-priced objectives.

Background

The paper discusses costly-observation bandits that provide cost-sensitive guarantees, including two-phase Gaussian-process algorithms with cost-adjusted information-gain complexities and matching finite-arm rates. It distinguishes those results from the paper's action-before-state selective-verification setting with a hard pathwise audit budget.

Within this comparison, the paper identifies general-kernel optimality as unresolved in the cited cost-sensitive Gaussian-process literature. The open problem concerns extending optimality results beyond finite-arm or otherwise restricted kernel settings while retaining observation-cost sensitivity.