... estimating¹

Since we cannot guarantee that all the pertinent data is available at training time, we can only produce an estimate of the accuracy of a classifier [28].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... problems²

Although we have concentrated on indexing undesirable system states, the same ideas can be used to capture ``favorable'' states.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.