Surrogate-Assisted Optimization

Surrogate-assisted optimization studies how to search efficiently when evaluating a candidate requires an expensive simulation, experiment, or oracle call. Common approaches include Gaussian-process Bayesian optimization, neural-network and ensemble surrogates, and surrogate-guided evolutionary search, often using uncertainty, gradients, trust regions, or multiple levels of fidelity to choose the next evaluations. The methods are used across engineering design, scientific discovery, and discrete or multi-objective problems, with newer work also applying generative models and learned optimization policies. A central challenge is that a model that predicts well on held-out data may still mislead an optimizer in regions it actively seeks, so reliable methods must manage uncertainty and distribution shift and validate promising candidates with the true evaluator.