surrogateassisted

Surrogateassisted, also written as surrogate-assisted, refers to a class of optimization methods that combine a computationally inexpensive surrogate model with an outer search strategy to optimize an expensive objective function. The surrogate provides fast approximations of the true objective, enabling more iterations of evaluation at a reduced computational cost. This approach is widely used in engineering design, materials discovery, hyperparameter tuning, and other areas where each evaluation of the objective is costly or time-consuming.

Common surrogate models include Gaussian processes (kriging), radial basis function networks, polynomial regression, support vector machines, and shallow neural networks. The outer search strategy can be a gradient-free optimizer or a global search method; Bayesian optimization is a prominent framework that uses a probabilistic surrogate to guide sampling via an acquisition function such as expected improvement, probability of improvement, or upper confidence bound. The typical workflow starts with an initial set of evaluations to train the surrogate, then iteratively selects new candidate solutions by maximizing the acquisition function, evaluates the true objective for a few promising candidates, and updates the surrogate with the new data. This loop balances exploration of unknown regions with exploitation of known good regions.

Advantages include reduced number of expensive evaluations and the ability to incorporate uncertainty. Limitations include potential model bias, sensitivity to the choice of surrogate and acquisition function, difficulties in high-dimensional or highly multimodal problems, and the need for careful design to avoid premature convergence. Surrogate-assisted methods are often combined with evolutionary algorithms, swarm intelligence, or gradient-free optimizers to handle noisy or discontinuous objectives.