Drift-Aware Preference-Based Bayesian Optimization with Dynamic Multinomial-Logit Surrogates A Confirmatory Semi-Synthetic Validation on NASA N-CMAPSS Maintenance Decisions

Preference-based Bayesian optimization (PBO) enables sequential optimization when candidate solutions can be compared but cannot be assigned reliable scalar values. Most PBO methods, however, assume stable decision-maker preferences. We introduce DA-MNL-PBO, a drift-aware framework in which interpretable multinomial-logit/Bradley-Terry preference coefficients evolve through a two-regime state-space model. A sequential Laplace approximation within an interacting multiple-model filter maintains stable and adaptive preference states, while a regime-aware acquisition rule combines optimization value with information about the preference coefficients and latent adaptation regime. We evaluate DA-MNL-PBO in a semi-synthetic predictive-maintenance study using NASA N-CMAPSS run-to-failure trajectories. Real engine trajectories and remaining-useful-life information define maintenance-policy outcomes, while controlled time-varying stakeholder utilities generate pairwise preferences. Hyperparameters were frozen before confirmatory evaluation. Across two confirmatory subsets, four temporal conditions, three response-consistency levels, 200 paired replicates, and four methods, the study comprised 19,200 method-level runs. Under preference drift, DA-MNL-PBO achieved lower mean cumulative dynamic regret than static and sliding-window MNL-PBO in all 12 confirmatory drift conditions, reducing regret relative to static MNL-PBO by 6.9-71.1% (42.1% on average). Fixed exponential discounting performed better under all abrupt shifts, whereas DA-MNL-PBO performed better in five of six gradual-drift conditions. The method also produced meaningful shifts in maintenance recommendations after preference changes. These results support probabilistic drift-aware preference modeling, particularly under gradual change, while demonstrating that fixed forgetting remains competitive under abrupt shifts. Because preferences are simulated, the results establish algorithmic adaptation rather than human preference validity.