Conformal Candidate Certification for Offline Model-Based Optimization

TMLR Paper9467 Authors

Abstract:

Offline model-based optimization (MBO) proposes candidates by optimizing
a surrogate trained on a fixed historical dataset.
Because candidates are deliberately out-of-distribution, surrogate rankings are
least reliable exactly where the optimizer is most aggressive.
Existing methods address this issue by regularizing the surrogate or
proposal mechanism, but they do not provide a per-candidate statistical
certificate that a proposed design meets a user-specified performance threshold.
We propose \emph{Conformal Candidate Certification} (CCC), a post-hoc wrapper that
attaches a calibrated one-sided lower bound to each candidate and
advances only those whose bound exceeds the target.
The key challenge is covariate shift: calibration data follow the
historical distribution, while candidates follow the proposal distribution.
We show that entropy-regularized surrogate maximization induces a
\emph{Gibbs-tilted proposal distribution}, allowing the same surrogate
that drives optimization to supply importance weights for weighted
conformal prediction, without a separate density-ratio estimation step.
Under oracle weights and strict data splitting, CCC satisfies
finite-sample marginal lower-bound validity.
Experiments on a synthetic stress test and the superconductivity dataset of (Hamidieh, 2018)
($n=17{,}011$ compounds) show selective certification with
empirical coverage at or above the nominal level, a $9.5$\,K gain in
mean certified critical temperature over the naive surrogate rule, and
a reduction in the false-acceptance rate by more than half.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Mohammad_Hajiesmaili1

Submission Number: 9467