New Bayesian Optimization Method Ensures Safety with Unobserved Baselines

Many decision-making processes, particularly in critical fields, require new interventions to maintain outcomes above a certain safety threshold, often compared to an existing standard. This research addresses a challenge in safe Bayesian optimization: how to ensure safety when the baseline policy's outcomes are counterfactual and thus unobserved. For instance, in clinical trials, a new treatment must not worsen patient outcomes compared to a standard of care whose exact counterfactual results for the new treatment group are unknown. The proposed solution involves using conformal prediction to construct valid uncertainty intervals for these unobserved counterfactual baseline outcomes. These intervals are then integrated into the safe Bayesian optimization process. This integration allows the system to maximize an objective while guaranteeing that safety constraint violations occur at or below a rate specified by the user. The paper also details how these conformal estimates can be adapted to handle different types of covariate shift, which is crucial for real-world applicability. A safety proof, experimental evidence, and a sensitivity analysis are provided to validate the method's robustness and effectiveness in ensuring safe optimization under uncertainty.