In complex engineering tasks—from training deep neural networks to designing bipedal robots—researchers face a "Goldilocks" problem. High-fidelity data, such as real-world robot trials or full-scale physical experiments, is accurate but prohibitively expensive. Conversely, low-fidelity data, like coarse simulations or short training runs, is cheap but often biased or noisy.
The research direction of Multi-Fidelity Bayesian Optimization (MFBO) provides a principled framework to navigate this tradeoff. Instead of treating all data points as equal, these methods treat simulations and physical experiments as different "information sources" with distinct price tags. By prioritizing information gain relative to cost, these algorithms can find optimal solutions with a fraction of the traditional experimental budget.
The Information Cost Tradeoff
The core intuition behind MFBO is that we should only pay for high-precision data when the model’s uncertainty cannot be resolved by cheaper alternatives. Alonso Marco, Felix Berkenkamp, and their colleagues Virtual vs. real: Trading off simulations and physical experiments in reinforcement learning with Bayesian optimization framed this as a resource allocation problem. They proposed that an optimizer should autonomously decide whether to query a cheap, noisy simulator or an expensive, precise physical experiment.
This decision is governed by a cost-sensitive acquisition function. While standard Bayesian Optimization (BO) might use Expected Improvement (EI) to find the next point to sample, MFBO methods like MF-GP-UCB, introduced by Kirthevasan Kandasamy in Gaussian Process Bandit Optimisation with Multi-fidelity Evaluations, use low-fidelity queries to "filter" the search space. This ensures that the expensive "high-fidelity" budget is reserved only for the most promising regions of the design space.
Normalizing Information by Expenditure
To make these decisions mathematically rigorous, researchers have developed metrics that normalize the value of information by its cost. Jialin Song and Yuxin Chen developed the MF-MI-Greedy framework, which calculates the ratio of mutual information gain to the computational or physical cost of the evaluation A General Framework for Multi-fidelity Bayesian Optimization with Gaussian Processes.
This "bang-for-your-buck" approach is particularly effective in materials discovery. Henry Herbol and Paulette Clancy demonstrated that using low-level simulations to guide high-fidelity calculations can reduce discovery costs by up to 80% Cost-effective materials discovery: Bayesian optimization across multiple information sources. Their PearsonKG algorithm selects the next sampling point by maximizing the expected value of information relative to the query cost, allowing the system to strategically switch between sources.
Modeling the Reality Gap
For MFBO to work, the model must understand how different fidelities relate to one another. Most frameworks use Gaussian Processes (GPs) to model these relationships. Alexey Zaytsev and Evgeny Burnaev provided a theoretical foundation for this, showing how to balance budget allocation between two sources to minimize interpolation error Minimax Error of Interpolation and Optimal Design of Experiments for Variable Fidelity Data.
In more complex scenarios, such as fluid dynamics, the output might be high-dimensional. Zheng Wang and Robert Michael Kirby developed MFHoGP, which uses a deep matrix structure to combine a few expensive samples with millions of low-accuracy data points Multi-Fidelity High-Order Gaussian Processes for Physical Simulation. This allows the model to learn the "shape" of the function from cheap data while using expensive data to correct the bias.
Beyond Myopic Decisions
A significant challenge in budgeted optimization is that the "best" next step might not be the most informative one for the entire remaining budget. Raul Astudillo and Peter I. Frazier addressed this with B-MS-EI, a non-myopic acquisition function that plans several steps ahead Multi-Step Budgeted Bayesian Optimization with Unknown Evaluation Costs. Similarly, Quan Nguyen and Roman Garnett introduced MF-ENS, which looks across the entire remaining budget to balance exploration and discovery in active search tasks Nonmyopic Multifidelity Active Search.
The Curvature of Cost
Why does redundancy sometimes beat precision? Mehmet A. Donmez and Andrew C. Singer explored the fundamental limits of this tradeoff in Cost-Performance Tradeoffs in Fusing Unreliable Computational Units. They found that the optimal strategy depends on the "curvature" of the cost-fidelity function. If increasing precision yields diminishing returns (a convex cost function), it is more cost-effective to "fuse" many cheap, noisy observations than to buy one expensive, high-precision measurement.
This principle is now being applied to modern AI challenges. For instance, Md Shahriar Iqbal and Pooyan Jamshidi developed FlexiBO, which decouples objective evaluations in Deep Neural Network design to prioritize experiments with the highest expected gain per unit of wall-clock time FlexiBO: A Decoupled Cost-Aware Multi-objective Optimization Approach for Deep Neural Networks. By recognizing that not all performance metrics cost the same to measure, they achieved significant improvements over standard, cost-blind methods.