New Bayesian Optimization Method Leverages Historical High-Fidelity Data

Black-box optimization, crucial in science and engineering, often involves expensive objective functions that can be approximated by cheaper, lower-fidelity proxies. Multi-fidelity Bayesian optimization (MF-BO) is a standard technique that exploits correlations between these different fidelities to efficiently query the objective. However, a significant challenge arises when the true highest-fidelity function is too costly to be included directly within the optimization loop. Many real-world scenarios, such as molecular optimization where top candidates are simulated before their true values are revealed, generate valuable "gold standard" high-fidelity data from prior experiments. Existing MF-BO algorithms often fail to optimally leverage this pre-existing, out-of-the-loop high-fidelity information. This work proposes a novel approach to mitigate this suboptimality by integrating historical high-fidelity data, along with task descriptors, into the MF-BO framework. The effectiveness of this new method has been demonstrated across synthetic functions and practical applications in chemistry and hyperparameter optimization.