Multi Fidelity Bayesian Optimization

Multi-fidelity Bayesian optimization (MFBO) accelerates the optimization of computationally expensive functions by strategically incorporating information from multiple sources of varying fidelity (accuracy and cost). Current research emphasizes efficient acquisition functions, adaptive fidelity selection strategies (like identifying "efficient points" and "saturation points"), and robust handling of heterogeneous errors and unreliable low-fidelity data, often employing Gaussian processes or deep Gaussian processes as surrogate models. MFBO's impact spans diverse fields, significantly improving efficiency in tasks such as materials discovery, controller falsification, hyperparameter optimization, and reinforcement learning by reducing the reliance on expensive high-fidelity evaluations.