Multi-Fidelity Optimization

It studies how to combine cheap, approximate evaluations—such as short training runs, coarse simulations, or small data subsets—with costly, accurate tests under limited budgets. Common approaches include multi-fidelity Bayesian optimization with Gaussian-process or transfer-learning surrogates, bandit and successive-halving resource allocation, and evolutionary search that screens candidates before high-fidelity validation. Applications range from hyperparameter tuning and materials or engineering design to model and system configuration. A central challenge is deciding when a proxy is trustworthy: bias, noise, and changing candidate rankings can mislead search, so methods adapt evaluation budgets and validate promising candidates while balancing cost and reliability.