Computational Intelligence

Computational Intelligence versus Artificial Intelligence 35

any values of x

∗

∈

/

S for which we know the correct f(x

∗

) . So we use some of the

dataset for training and some for testing.

We usually assume that the ability of a model to generalize is best measured by

the system perfor mance on the test set. It is quite possible that the best test set per-

formance does not coincide with the best performance on the training set. A neural

network, for example, can be overtrained on the training set (it is said to “memorize”

it) so that it performs relatively poorly on the test set.

In summary, it is important to define what you mean when you use the term

generalization ...