The use of surrogate models (also known as meta-models
or response surface models) to reduce the number of candidate
solution evaluations when they are computationally
expensive or difficult to obtain/formulate has been developed
as evolutionary computation has been applied to more
complex domains, e.g., in systems where a human user is
involved [30]. This is typically achieved through the construction
of models of the problem space via direct sampling—the
use of approximations is an established approach in the wider
field of optimization [31]. That is, the evolutionary process
uses one or more models to provide the (approximate) utility
of candidate solutions, thereby reducing the number of real
evaluations during iterations. Initially, all candidate solutions
must be evaluated directly on the task to provide rudimentary
training data for the modeling, e.g., by neural networks.
Periodically, high utility solutions suggested by the model
optimization are then evaluated by the real system. The training
data for the model is then augmented with these and the
model(s) updated. Over time, as the quality of the model(s)
improves, the need to perform real evaluations/fabrications
reduces.